Hyperspectral Imaging System Applications in Healthcare

Abstract

Hyperspectral imaging (HSI) is a swiftly developing intraoperative and diagnostic technique in several clinical specialties. By monitoring oxygenation and biochemical markers, it helps with tissue viability, burn depth measurement, wound healing, and tumor detection. HSI facilitates real-time, harmless diagnosis throughout surgeries or outpatient settings, and allows for the detection of tumor boundaries with over 90% accuracy, according to clinical studies. Originally developed for remote sensing and aerospace applications, HSI has rapidly evolved and found increasing relevance across diverse sectors, including agriculture, environmental monitoring, food safety, pharmaceuticals, defense, and especially medical diagnostics. This review explores the origins, development, and expanding applications of HSI, with a particular emphasis on its role in healthcare. It discusses the operational principles and unique features of hyperspectral systems, such as their ability to produce spectral data cubes, perform non-destructive analysis, and integrate with emerging technologies like artificial intelligence and drone-based platforms. By comparing hyperspectral imaging to traditional and multispectral techniques, the review highlights its superior spectral resolution and versatility. Key challenges, including data volume, sensor calibration, and real-time processing, are also addressed. Finally, emerging trends such as miniaturization, integration with the Internet of Things, and sustainable system designs are examined, offering insights into the future directions and interdisciplinary potentials of HSI in both scientific research and practical applications.

1. Introduction

Medical imaging has become an important part of modern healthcare services since it involves technologies that aim to diagnose and monitor patients [1]. Medical imaging technologies include X-rays, ultrasonography, mammography, computed tomography (CT scans), and other related tools [1]. The availability of these technologies has changed how abnormalities are diagnosed by providing visual representation that helps medical professionals become aware of their patients’ situations [2]. While visual information is the basis of medical imaging tools, some devices can also be considered medical imaging technologies due to their outputs. For example, electrocardiography (ECG) records and monitors certain activities in humans, but does not go as far as providing visual information on what is going on. However, the representation of data as a parameter graph can serve as a useful alternative [1].

Medical imaging tools date to the use of observation by ancient doctors to diagnose ailments and diseases [2]. However, this developed into the detection of diseases using saliva and body fluids. X-rays and microscopes were developed in the 19th century and served as the basic diagnostic tools for illnesses. More advanced techniques were developed after the development of X-rays, including fluoroscopy, which was converted into computed tomography. Similarly, the X-ray beam was used to develop mammography to generate high-resolution breast images. In recent years, there have been computerized axial tomography devices that serve more sophisticated functions compared to the initial tomography [3]. Despite the worldwide application of conventional modalities such as MRI, CT, and X-rays, there is a gap in the visualization of structural or morphological data or lack of functional or biochemical insights [4].

With advancements in geophysics and information communication technologies, HSI offers an opportunity for advancement in the medical imaging field [5]. HSI is a technique that provides information about unique objects in the human body. It works by collecting and processing information from different parts of the electro-magnetic spectrum and identifying materials, finding objects, or detecting a process. Three hyperspectral imagers may exist within the broad hyperspectral imaging (HSI) category, including push broom scanners, which read images over time; band sequential scanners, which acquire images at different wavelengths; and snapshot hyperspectral imagers, which use an array to generate the image of what is being observed instantly.

One of the features of HSI is that more bands of light can be perceived than by the normal human eye. As a result, a wider range of wavelengths can be recorded using HSI. For this purpose, the application of the HSI technique extends to various industries to understand and implement systems that can help drive an understanding of what is being observed in astronomy, molecular biology, geosciences, physics, medical sciences, and other fields. Additionally, in the field of healthcare, even when tissue types look the same in visible light, HSI allows for their identification based on distinct spectral fingerprints. This involves using high spectral sensitivity to analyze oxygenation, hemoglobin distribution, and metabolic makeup in order to distinguish malignancies from healthy tissues [6].

In neurological setups, where conventional modalities like MRI could miss metabolic changes, HSI has shown useful in detecting early-stage brain disorders. Alteration in mitochondrial activity in malignancies have been connected to hyperspectral indicators like cytochrome oxidase absorption [3].

Furthermore, by recording physiological indicators such as the volume of blood, oxygenation, and tissue perfusion, HSI provides practical insights in contrast to static anatomical imaging. In surgical settings, these are crucial for assessing ischemia, flap viability, and the healing of wounds [4].

HSI has been significant in ensuring that there is an opportunity for researchers and government authorities to make informed decisions. For example, the analysis conducted by Akewar et al. [7] demonstrated that HSI ensures that professionals can assess relevant aspects of their works through what is being observed using HSI [7]. Thus, informed decisions are achieved through the combination of spectral and spatial data that facilitates in-depth analysis across several disciplines.

1.1. Objectives of the Review

The objective of this research is to summarize and demonstrate the clinical uses of hyperspectral imaging (HSI) in the medical field. It specifically examines the use of HSI in burn wound evaluation, gastrointestinal imaging, dermatological diagnosis, tumor identification, surgical guidance, and perfusion monitoring. This research review intends to review the application of this technique by considering its relevance and its impact within the health field.

As a result, several themes are explored in this review. These themes include, but are not limited to, the following:

• Analysis of the relevance of HSI in the context of contemporary healthcare;
• The exploration of the unique features of HSI, including the sensors, calibration, and types of images that may be produced using the HSI technique;
• Comparison of HSI with other techniques to determine the strengths and limitations of HSI systems;
• Recommendations on the process to improve HSI by considering the opportunities from technological breakthroughs.

1.2. Selected Sources

Multiple sources were used to complete this study review. The following

Table 1. Comparison table summarizing main sources.

Author(s), Year	Method	Application	Findings
Hussain et al., 2022 [8]	Diagnostic Imaging	Medical Field	Disease detection through modern imaging.
Kumar et al., 2021 [9]	Comparative Study	Medical Imaging	Image modality comparison
Haleem et al., 2022 [10]	Medical 4.0 Tech	Healthcare	IoT and cyber-physical systems
Oliveira Neto et al., 2024 [11]	Saliva Diagnostics	Viral Infections	Diagnostic tool potential
Karim et al., 2023 [12]	HSI Overview	Medical Imaging	Trends in hyperspectral imaging
Wu et al., 2024 [13]	Diagnostic Imaging	Hyperspectral images	Head and Neck cancer results accuracy
Akewar & Chandak, 2024 [7]	HSI Algorithms	General Imaging	Challenges include data complexity and algorithm limitations.
Perri et al., 2024 [14]	FT HSI Camera	Optical Sensing	Fourier-transform hyperspectral camera
Magnusson et al., 2020 [15]	Color Matching	Hyperspectral	RGB creation from HSI data
Wan et al., 2021 [16]	Remote Sensing	Ore Investigation	HSI in mining exploration
Qian, 2021 [17]	Hyperspectral Satellites	Remote Sensing	Evolution of hyperspectral satellites
Lu & Fei, 2014 [18]	HSI Overview	Medical Imaging	Medical HSI applications
ESA, 2023 [19]	HSI Overview	General	Space-based HSI applications
Innoter, 2023 [20]	HSI Overview	General	HSI tech applications in everyday life
Nireos, 2023 [21]	HSI Overview	General	Applications in multiple sectors
JOUAV, 2024 [22]	HSI Overview	General	Applications in surveillance & security
Morales et al., 2022 [23]	HSI System Setup	Laboratory	Hyperspectral system validation
Kushalatha & Prasantha, 2022 [24]	HSI Preprocessing	General	Pre-processing techniques for HSI
Shaikh et al., 2021 [25]	HSI Calibration	Imaging Systems	Low-cost calibration method
Sellar & Boreman, 2005 [26]	Imaging Classification	Remote Sensing	Classification methods for remote sensing
Geladi et al., 2004 [27]	HSI Calibration	General	Calibration challenges in HSI
Hruska et al., 2012 [28]	HSI Analysis	UAV Imaging	UAV-based HSI analysis for remote sensing
Mazdeyasna et al., 2025 [29]	Normalization Methods	HSI Cameras	Normalization for spectral performance
Zhang & Abdulla, 2023 [30]	CNN and Batch Normalization	HSI Classification	Optimizing classification performance
Wang & Chang, 2006 [31]	ICA Dimensionality Reduction	HSI Analysis	ICA for dimensionality reduction in HSI
Bilgin et al., 2008 [32]	Unsupervised Classification	Remote Sensing	Fuzzy-based classification of HSI
Lv & Wang, 2020 [33]	HSI Overview	Image Classification	HSI classification methods reviewed
Zhao & Du, 2016 [34]	Feature Extraction	HSI Classification	Spectral-Spatial feature extraction
Hu et al., 2022 [35]	Anomaly Detection	Remote Sensing	Deep learning for anomaly detection
Cui et al., 2022 [36]	Deep Learning	Medical HSI	Applications of deep learning in medical HSI
Booysen et al., 2022 [37]	HSI of Minerals	Remote Sensing	HSI for mineral detection in Africa
Zhang et al., 2020 [38]	Tumor Detection	Medical	HSI in solid tumor diagnosis
Datta et al., 2022 [39]	HSI Classification	General	Challenges in HSI classification
Sucher et al., 2024 [40]	HSI Analysis	Medical Imaging	HSI enabled intraoperative perfusion monitoring
Markgraf et al., 2020 [41]	Anomaly Detection	Medical	Accurate prediction of kidney tissue
Liu et al., 2025 [42]	Comparative study	Medical Imaging	Image modality trends
Bhargava et al., 2024 [43]	Disease Detection	Medical Imaging	Challenges in clinical translation
Lai et al., 2024 [44]	HSI Application	Medical Imaging	Disease detection through medical imaging
Kim et al., 2017 [45]	Diagnostic imaging	Healthcare/Cardiovascular	Smartphone-endoscope with LSCI enabled real-time blood perfusion mapping
Boese et al., 2022 [46]	Comparative Study	Medical Imaging	Image modality comparison
Hsu et al., 2017 [47]	HSI Algorithms	Healthcare	clustering-based compression improved telemedicine data transfer and storage
Luo et al., 2022 [48]	Medical 4.0 Tech	Medical Imaging	AI-enabled edge–cloud telemedicine
Zhang et al., 2024 [49]	HSI Overview	General	Trends in hyperspectral imaging
Hazarika et al., 2024 [50]	Saliva Diagnostics	Healthcare	Early-stage development of non-invasive biosensors for neonatal jaundice
Nanegrungsunk et al., 2022 [51]	Diagnostic Imaging	Medical Imaging	Review of retinal imaging techniques in diabetic retinopathy
Dietrich et al., 2021 [52]	HSI Analysis	Medical Field	HSI monitored microcirculatory oxygenation and perfusion
Bhatti et al., 2025 [53]	Deep Learning	Detection and delineation of glioma tissue	Deep learning enhances HSI healthcare accuracy.
Weber et al., 2025 [54]	FLAIR-MRI sequence	Develops DL algorithms for endoscopic HSI data	Accuracy in detecting non-enhancing glioma tissue Intraoperative oral diagnostics and noninvasive pathological tissue analysis for early cancer detection.
Romer et al., 2025 [55]	Endoscopic HSI system	Colorectal cancer detection	Provide unique preprocessing configurations for better classification performance in medical imaging.
Tkachenko et al., 2025 [56]	Spectrum scaling	Cancer diagnosis	Enable improved classification compared to raw spectral features.

is a comparison table summarizing the sources used in this study.

2. HSI and Related Concepts

Hyperspectral imaging involves acquiring images at multiple wavelengths across the entire electromagnetic spectrum, including visible light, the infrared, and ultraviolet ranges. Each pixel in a hyperspectral image contains a full spectrum of information, which forms a hyperspectral cube. The cube is made up of two spatial dimensions and one spectral dimension, which allows for a complete analysis of the imaged scene. The hypercube is typically a stack of images of the same object or scene. The ability to analyze light in such detail makes it possible to identify materials based on their specific reflectance characteristics. The technology behind hyperspectral imaging involves effective sensors that can measure and record the intensity of light reflected by objects across these many wavelengths. These high-resolution data make it possible to differentiate between materials that have similar appearances but with different chemical compositions [14].

2.1. Distinction from Multispectral Imaging

Both multispectral imaging (MSI) and hyperspectral imaging (HSI) have been thoroughly studied in the medical field, especially in the areas of diagnostics, surgical guidance, and digital pathology. Extensive spectral patterns that allow for accurate tissue differentiation are obtained by HSI, as opposed to MSI, which usually only records a few broad spectral bands. It has been demonstrated that this increased spectral resolution improves diagnostic capabilities. Compared to traditional RGB imaging, HSI/MSI has been used in digital and computational pathology for tasks including digital staining, color adjustment, and immunohistochemistry, enhancing disease identification throughout skin, gastrointestinal, and kidney tissues [13,21]. Similarly, in dermatology, HSI and MSI help differentiate benign nevi from melanomas, which is a crucial non-invasive diagnostic capacity. Newer studies are even incorporating these systems into handheld devices.

HSI’s comprehensive spectral information enables the real-time monitoring of oxygenation in tissues, vascular perfusion, and tumor borders in surgery and intraoperative settings. In complex medical contexts where accurate tissue identification can save lives, HSI’s narrow-band spectra features enable better discrimination and detection, while MSI provides faster data capture and easier integration [57].

2.2. Comparison with Traditional Imaging

Traditional imaging methods use RGB channels to represent normal colors. Every pixel is defined based on its intensity value. On the other hand, HSI maps a full spectrum to each pixel, which provides more information than just the simple color representation [16]. This capability enables accurate and detailed analysis. Analysis is based on material properties in terms of how they reflect and absorb light at different wavelengths. HSI is also able to capture continuous spectral data across multiple narrow bands. This helps provide detailed information about material characteristics that traditional and multispectral imaging cannot provides [58]. Keeping in mind the role of the study that is inclined toward the role of HSI in healthcare, hyperspectral imaging (HSI) provides extensive biochemical as well as physiological insights that are crucial for medical diagnosis and surgical guidance since it maps entire spectrum data to each pixel rather than just RGB values. For example, intraoperative HSI has demonstrated the ability to distinguish between healthy cartilage and degenerated tissue during knee surgery in orthopedics by measuring spectral absorption differences, particularly at 540 nm, with a sensitivity for diagnosis and specificity of 81 percent, which is comparable to magnetic resonance imaging (MRI) [59].

To conclude, by taking advantage of the material-specific light absorption and reflectance characteristics of living tissues, HSI’s continuous, narrow-band spectral mapping improves surgical results and diagnostic precision in comparison to traditional imaging that does not have a high level of precision.

2.3. The Electromagnetic Spectrum and HSI

Electromagnetism forms the basis of hyperspectral imaging (HSI), which captures how materials interact with light across a broad range of the electromagnetic spectrum. In HSI, each pixel contains spectral information from multiple contiguous bands, allowing for the precise detection of biochemical and structural changes in tissues. This makes it a powerful tool in medical diagnostics and intraoperative imaging.

By extending its spectral range from the visible to near infrared (NIR), hyperspectral imaging (HSI) has made significant advances in the field of medical imaging. Systems currently operate between 490 and 1600 nm. A well-researched study from 2024 presented a rigid endoscope that combines an acousto-optic adjustable filter and a supercontinuum light source to capture real-time HSI during surgery, effectively discriminating nerves, blood arteries, and malignant tissue with great spectral fidelity. This advancement is a prime example of how machine learning-based tissue classification is supported by extended-spectrum HSI, allowing for safer and more accurate intraoperative judgments. In addition, 400–1000 nm HSI with 3D convolutional networks was used in a newly validated head-and-neck surgical investigation to automatically recognize glandular, nerve, muscle, and vascular components, with accuracies above 83% [13].

Additionally, an innovative development that was published in the Journal of Medical Internet Research in 2025 showed how to detect cardiac failure via preserved ejection fraction (HFpEF) non-invasively through oral HSI using 25 spectral bands and random forest algorithms. The application reported diagnostic accuracies ranging from 83 to 86 percent on internal as well as external patient cohorts. These results highlight how HSI, which spans visible and NIR wavelengths, may be used in cardiovascular medicine for a label-free, non-invasive detection technique. Together, these innovations demonstrate how HSI may use tissue-specific light absorption and reflectance to guide surgeries in real time and identify diseases, underscoring its revolutionary significance in modern medical imaging [57].

3. Fundamentals of Hyperspectral Imaging in Biomedical Context

3.1. HSI Data Cube

A three-dimensional dataset is produced by his, commonly defined as a hypercube. It has two dimensions (x, y) that reflect the spatial information of the tissue. The third dimension (λ) indicates the spectral domain. When compared to conventional RGB imaging that covers three broad wavelength bands, HSI covers hundreds of spectral bands for each pixel and assigns a unique spectral fingerprint unique to its structural and biochemical composition. Researchers access spectral patterns and spatial properties of tissue simultaneously to support biochemical and morphological characterization [60].

3.2. Acquisition Methods Through Hypercube

The hypercube can be acquired using different scanning strategies, e.g., spatial scanning through whisk broom based on point scanning and push broom that entails line scanning will be discussed in detail in Section 6. The stage enables high spectral resolution but remains sensitive to motion artifacts. The spectral scanning method captures a 2D image at one wavelength and repeats the process to develop a cube through successive wavelengths. This method keeps spatial field of view intact but subject motion can produce distortions. The snapshot imaging is another fundamental concept of HSI that uses prism, coded apertures, and micro-filter arrays to capture the entire cube in a single exposure. The method is suitable for in vivo monitoring and surgery due to real-time acquisition but compromises spectral resolution or spatial coverage [18].

3.3. Spectral Basis in Tissue

The applicability of HSI in healthcare is evidenced by the fact that tissues contain chromophores that absorb light at specific wavelengths, e.g., lipids, deoxyhemoglobin, water, oxyhemoglobin, and melanin. The difference in their relative concentrations has the ability to change the spectral signature of tissue. This relationship is called the Beer–Lambert law as it provides a linkage between absorbance (A) at a given wavelength to chromophore concentration (c), path length (l), and molar absorptivity:

A(λ) = ε(λ)⋅c⋅l

It can help indicate variations in hydration, metabolic activity, and tissue oxygenation non-invasively through spectral changes in hyperspectral data [61].

3.4. Spectral Ranges in Medicine

In biomedical imaging, variation in spectral ranges provides additional information. Visible spectral range between 400 and 700 nm is sensitive to hemoglobin absorption, and is useful for analyzing vascularization and blood oxygenation. The near-infrared spectral range is between 700 and 1000 nm, and holds deeper tissue penetration. However, it is sensitive to water and lipid content, but applied rigorously in wound assessment and tumor margin detection. The short-wave infrared spectral range is between 1000 and 2500 nm with strong absorption probability by lipids, proteins, and water. It is also used in burn depth analysis and metabolic imaging. The mid-infrared range is >2500 nm with limited penetration depth, and provides molecular-level vibrational signatures, though less in vivo monitoring. The appropriate spectral range enables the non-invasive mapping of tissue states through HSI in a wide range of clinical applications [1].

4. Historical Development

4.1. Early Development of HSI

In order to evaluate the Earth’s surface and atmospheric conditions using continuous spectra in hundreds of tiny bands, NASA launched aerial imaging spectrometers like AIS and AVIRIS in the mid-1980s, which is when hyperspectral imaging first emerged [59]. This was followed by the development of hyperspectral technology in Canada in the 1980s–1990s. To add further, the medical sector used HSI technology to understand the underlying cause and presence of diseases. Therefore, the adoption of HSI technology can be traced back to a long history of adopting various observation techniques to measure and diagnose diseases [62]. In the early times, hyperspectral systems (remote-sensing cameras that weighed hundreds of kilograms and cost millions of dollars) were only used in geological and aerospace applications due to their high cost and mass, but advancements with the passage of time resulted in great usage of HSI in the medical field. Wider adoption was made possible by sensor miniaturization, enhanced optical components, and increased computer power during the 1990s and early 2000s. By the 2010s, these innovations enabled HSI to move from massive aerospace components to ground-based and, ultimately, portable versions [63].

4.2. Historical Background of HSI in Healthcare

Beginning with microscopy to examine chemical composition at the cellular and subcellular levels, HSI joined the biological research field between 2010 and 2020. Applied spectrum imaging was established in 1993 and introduced spectrum microscopes for automated pathology and cytogenetics. Later, to describe infections and nanoparticles in biological preparations, CytoViva combined dark-field microscopy combined with nanoscale hyperspectral capture in the healthcare sector [62].

Around 2020, the clinical use of hyperspectral imaging (HSI) for healthcare purposes increased dramatically, primarily as a result of breakthroughs in push-broom and snapshot camera technologies that made it possible to conduct proof-of-concept studies in a variety of medical specialties. By distinguishing spectral signatures among malignant and healthy tissues, HSI systems in oncology showed great accuracy and precision in identifying colorectal and brain tumors during surgical procedures. These features represented a significant advancement in image-guided, real-time interventions. Moving further, the “Hyper scope,” a portable dermatoscope that can record spectra in the 400–950 nm range, was created as a result of advancements in dermatology and is ideal for non-invasive outpatient skin examinations. Additionally, HSI has shown effectiveness in oxygen monitoring and wound assessment, where it makes it easier to create oxygenation heatmaps that help treat illnesses like diabetic foot ulcers and Raynaud’s phenomenon. In another event, to increase its diagnostic potential, HSI has been studied in ophthalmology, specifically in retinal imaging, where scientists are looking at how it can identify β-amyloid deposits in the retina non-invasively to find early indicators of Alzheimer’s disease. These advancements collectively show how HSI is evolving into a more adaptable instrument across clinical and healthcare field, improving early detection and facilitating more accurate, individualized treatment plans in the field of medicine [63].

4.3. Technological Evolution of HSI Systems in the Medical Field

Hyperspectral imaging (HSI) healthcare-oriented significance has been greatly enhanced by technological advancements, especially in the areas of sensor hardware, AI integration, and portability. With a classification accuracy of up to 91%, advanced sensor technologies like push-broom fusion, which combines visible near-infrared (VNIR) and near-infrared (NIR) spectral bands between 400 and 1700 nm, have significantly improved tumor delineation in brain surgery. In a similar vein, real-time intraoperative imaging has been made possible by the launch of snapshot cameras such as the Cubert UHD 185 and Imec’s tiny sensors, which provide surgeons with access to spectral data while performing surgery [62]. In addition to these advancements, new mid-infrared scan-free imaging methods may now record complete spectrum profiles in a matter of seconds, opening up possibilities for molecular fingerprinting in fast-turnaround diagnostics. The diagnostic features of HSI have been substantially improved by the incorporation of machine learning and artificial intelligence. Machine learning-based tissue classification techniques have been demonstrated to cut processing times for glioblastoma surgery to less than three seconds, while neural networks using platforms like TIVITA have reached 86% sensitivity and 95% specificity in the identification of colorectal cancer. Furthermore, the spectrum examination of exosomes using Fourier-transform infrared spectroscopy (FTIR) and machine learning presents encouraging prospects for non-invasive early cancer screening. For example, exosome FTIR + SVM classifier for breast cancer showed 92% accuracy, 88% sensitivity, and 90% specificity [21]. Additionally, with gadgets like the hyperscope dermatoscope, which can record spectra between 400 and 950 nm, that enables outpatient skin diagnostics, the mobility and clinical utilization of this technology has also increased. Real-time guiding during neurosurgery procedures and wound management is being provided via handheld and augmented reality-enhanced surgery instruments that incorporate snapshot HSI. Additionally, wearable HSI tools for vascular oxygenation, dental diagnostics, and ongoing skin lesion monitoring are receiving more and more attention in research, bringing the technology much nearer point-of-care settings [43].

Strong clinical results emerged from these developments, highlighting the models where stomach cancer has accuracy rates above 94%, while models for detecting skin cancer have accuracy rates between 84% and 95%. As a result, the demand for cancer diagnoses, surgical guiding, and mobile diagnostic solutions is expected to cause the global HSI market, which was valued at around USD 1.6 billion in 2022, to rise at a compound yearly growth rate of roughly 13% to reach over USD 4 billion by 2030 [63]. When taken as a whole, these innovations are not only broadening the scope of technology but also moving us closer to a time when routine medical care will include safer, quicker, and more accurate diagnosis [23].

5. HSI Systems and Technologies

5.1. Components of the HSI System

5.1.1. High Spectral Resolution

One of the main characteristics of HSI systems is the ability to capture data in many narrow spectral bands. This allows HSI systems to detect even the slightest differences in materials, which makes them useful for precise identification and classification. In contrast, multispectral images capture data in a few broad bands [64].

5.1.2. Creation of Hyperspectral Data Cubes

HSI systems generate a hyperspectral data cube in which two dimensions represent spatial information and the third dimension represents spectral information. Each pixel in this cube contains a complete spectrum of light for the scene. This allows for the detailed analysis of the chemical and physical properties of a sample [24]. The three-dimensional representation allows for the detailed visualization and analysis of the data.

5.1.3. Versatile Applications

The versatility of HSI systems is demonstrated by their many applications in different fields. In agriculture, HSI can be used to monitor crop health [49]. In medicine, it helps improve diagnosis and enables image-guided surgeries.

5.1.4. Non-Destructive Analysis

A major benefit of HSI is that it is non-destructive. It can analyze samples without damaging them. Thus, it crucial in fields like cultural heritage conservation and medical diagnostics [25]. This capability allows for the preservation of valuable artefacts while still obtaining necessary analytical data.

5.2. Types of Hyperspectral Sensors

5.2.1. Push-Broom (Line Scanning) Sensors

Push-broom sensors capture spectral data line. It involves a linear array of detectors working simultaneously to record reflectance. As the sensor, lines of data are collected. This results in a hyperspectral image.

5.2.2. Whisk-Broom (Point Scanning) Sensors

Whisk-broom sensors work by sweeping back and forth on the surface to collect spectral data. The sensors have a smaller field of view and capture reflections only from specific points and not in continuous lines. The data are collected sequentially, leading to longer acquisition times than in the case of push broom systems. They are preferred where detailed information is required [65].

5.2.3. Snapshot Imaging Sensors

Snapshot imaging sensors capture an entire spectral image in one shot. This capability enables fast data acquisition without any scanning movements. They are ideal for changing situations where fast and continuous analysis is required. This is why they are widely used for medical diagnostics [66].

5.3. Platforms for HSI Data Acquisition

Airborne and satellite platforms are popular systems for hyperspectral data acquisition. They use advanced sensors that capture large spectral data covering vast geographic areas. For instance, the airborne visible and infrared imaging spectrometer (AVIRIS) of NSA is useful in environmental monitoring. The platforms cover large areas quicky, which make them suitable for monitoring vegetation health, assessing water quality, and detecting pollutants [67].

Ground-based HSI systems are for localized studies. These systems can be attached on tripods and used to capture high-resolution spectral data. They have a hyperspectral camera combined with a motorized stage that facilitates accurate movement across the target area [68]. They are useful in agricultural applications since detailed analysis is required.

Handheld and portable hyperspectral imagers are used in on-site analysis and do not need extensive setup, and hence, there is no need for specialized training. Open-source hyperspectral imagers provide a customizable solution since they can be employed remotely and attached to mobile platforms like drones [69]. This allows for rapid data collection in various environments. The combination of HSI and drones has enabled precision in large projects that require accurate data collection like in agriculture.

5.4. Advances in Sensor Technology

HSI has made great strides in healthcare applications where it has helped improve resolution, speed, and portability. It is also used in evaluating the quality of red ginseng by monitoring indicators such as reducing sugar content, water content, and hollow rate. Together with machine learning algorithms, HSI now provides fast and efficient monitoring, and helps reduce costs [29].

6. Pre-Processing and HSI Systems in Healthcare

The process of data acquisition and pre-processing is carried out to ensure quality, consistency, and reliability of data. Pre-processing steps included white-reference calibration and dark-current subtraction, spectral smoothing through noise reduction, spectral band selection based on wavelengths, background masking, and image registration. The dataset has to be normalized and standardized to maintain consistency across spectral bands [30].

6.1. Acquisition of Reflectance and Absorbance

In healthcare, HSI pre-processing involves data normalization and image registration. Normalization involves converting hyperspectral radiance observations to reflectance or absorbance values [70]. This is carried out to minimize system noise and eliminate image artefacts. Normalization techniques like reflectance and absorbance conversion are also employed to achieve this. Reflectance may be achieved by comparing raw data to a standard reference. The other method is geometric correction, which helps ensure proper spatial alignment [34]. On the other hand, absorbance is calculated from the ratio of sample to reference images. While hyperspectral imaging can provide live spectral images, image registration remains essential even in stationary settings, because the sequential acquisition of spectral bands can still introduce misalignments due to sensor variability, illumination changes, or subtle tissue motion. In dynamic or in vivo applications, live imaging further helps correct for patient or organ motion [35]. These pre-processing steps ensure that the quality and reliability of HSI data are achieved. This, in turn, enables a more accurate feature extraction, dimensionality reduction, as well as analysis in medical diagnostics, among other applications.

6.2. Image Acquisition

6.2.1. Methods for Image Acquisition

HSI plays an important role in diagnostic and monitoring applications. It provides detailed spatial and spectral data [37]. The method of image acquisition used may affect the accuracy of capturing spatial and spectral information. In this regard, HSI systems are classified based on their acquisition mode.

The first classification consist of the spatial scanning mode for image acquisition. It involves capturing a complete spectrum for each pixel location across whole image. Images can be acquired through two methods. Whisk-broom (point-scanning) systems acquire their data pixel by pixel. On the contrary, push-broom (line-scanning) systems capture data line by line because they scan across a scene in a spatial way. Even though these methods are useful for detailed imaging, they do not provide live spectral images because the hyperspectral data cube is built only after the full spatial scan is completed.

The other classification of HSI systems is the spectral scanning mode for image acquisition. The system scans through the wavelengths (spectral bands) while capturing the entire image at each band. It captures the entire scene in a single exposure using two-dimensional detector arrays. The system then goes through the various wavelengths to construct the hyperspectral data cube. This approach is beneficial for healthcare applications as it allows for the displaying of live spectral images [68]. This also enables real-time visualizations. It is advantageous in stationary settings like imaging biological tissues, as well as samples under a hyperspectral microscope, where precise targeting and focusing are required. A major advancement in healthcare-focused HSI is Fourier-transform infrared imaging (FTIR). FTIR systems combine a Fourier-transform spectrometer with a focal-plane array (FPA).

6.2.2. Normalization and Scatter/Glare Correction

The applications of standard normal variate (SNV) and multiplicative scatter correction (MSC) algorithms can help minimize the effect of surface reflections and scatterings. SNV subtracts the per-spectrum mean and divides it by standard deviation, while MSC fits individual spectra to a reference to provide accurate multiplicative and additive effects. Some researchers also employ min–max normalization, single-wavelength normalization or area under the curve (AUC) normalization. The application depends on the context and diagnostic relevance of optical contrast (e.g., blood absorption vs. scattering) [71].

6.3. Noise Reduction and Correction

Noise can affect the quality and application data. Reliable analysis is ensured through advanced noise reduction techniques that help correct for the effects of noise sources [34].

6.3.1. Sources of Noise

The major sources of noise in hyperspectral data are instrumental and environmental factors. Sensor-related issues like thermal noise, quantization noise, and photon noise are also a problem [38].

6.3.2. Noise Reduction Techniques

The major techniques used to address noise in hyperspectral imagery include smoothing filters, spectral denoising. Advanced techniques such as Gaussian filters are utilized to reduce variations arising at the pixel level. Moreover, Savitzky–Golay filtering is also another technique that is used to smoothen spectral bands white maintaining peak features that are significant medical environments where such features define pathological changes. Wavelet-based denoising is also utilized to decompose spectral signals into multi-resolution scales. The technique retains accurate biomedical signatures and also suppresses random noise. Non-local meets global (NGMeet) is another advanced denoising technique to examine tumor images and clinical wounds and maintain chromophore signatures in high-dimensional data [72].

Noise reduction techniques are sometimes classified into 3D model-based/3D filtering, penalty-based approaches, low-rank techniques, and mixed-noise models. Three-dimensional model-based/3D filtering technique relies on spectral and spatial structure jointly such as interpolated block-matching and 3D filtering and guided filtering (IBM3DGF). The denoising method involves obtaining inter-spectral correlation coefficients that divides HSIs into the groups to produce HSIs by utilizing three images to interpolate and conduct block-matching and 3D filtering (BM3D) to reduce the noise level of each group and inverse interpolation is applied to retrieve HSI [73].

Similarly, the penalty-based approaches utilize spectral/spatial smoothness penalties such as Laplacian or TV regularization for ensuring spectral integrity. The low-ranking methods are based on matrix/tensor decomposition like Robust PCA, Tucker decomposition for the purpose of isolating noise since it has a high probability in face of low-rank spectral signals. Mixed-noise models consider Gaussian, impulse, and striping presence in order to reduce noise with hybrid optimization frameworks [74].

6.3.3. Smoothing, Band Pruning, and Hybrid Methods

The smoothing filters such as Savitzky–Golay can be used to preserve absorption peak shape and reduce high-frequency spectral noise. In HSI-based medical research, such filters are used to distinguish chromophores. Additionally, band pruning can be used for removing low SNR bands or low signal. It also addresses bands that hold water absorption features. To reduce the effects of mixed noise or structured noise like impulse or stripe noise, hybrid methods that involve low-rank spectral decomposition and spatial smoothing along with neural denoising are also employed such as FastHyMix at the beginning to remove noise and preserve spectral details [75].

6.3.4. Noise at Spectral-Range Edges

HSI systems reflect elevated noise at the edges of spectral range, causing a drop in sensor sensitivity with lower signal-to-noise ratio. To reduce this noise, bands at the margins (e.g., <420 nm or >950 nm in VNIR systems) are mostly excluded during preprocessing stage for classification accuracy and data reliability [76].

6.4. Background Removal Through Region-of-Interest Segmentation

HSI datasets in clinical experiments hold irrelevant regions in a hyperspectral cube such as instrumentation artifacts, lighting fixtures and surgical drapes that cause dimensionality, and must be removed before analysis. The step involves the extraction of region of interest (ROI) manually, automatically through spectral thresholding and unsupervised clustering algorithms (mean shift and k-means) or by annotating boundaries of tissues using clinician-annotated tissue boundaries. Successively, a binary mask is used to separate and classify tumor and normal tissue song while maintaining computational efficiency [36].

6.5. Band Wavelength and Correlation

6.5.1. Selection of Wavelength Band

Sensors used in HSI gather insights on spectral data across contiguous bands. The extreme range of spectral features decrease signal quality due to imitation in the detector and optical distortion. The restriction of image analysis to relevant bands such as 900–1000 nm for water content or 520–850 nm for hemoglobin absorption enhances the accuracy of the model. Additionally, certain studies deploy statistical metrics for the automated selection of bands such as PCA or mutual information loading to identify the relevant wavelengths [77].

6.5.2. Correlation of Neighboring Spectral Bands

Neighboring bands in HSI data hold high correlation due to limitations in resolving the power of sensors and the gradual changes in spectral bands. It has varying computational cost and reduces diagnostic value. To address these issues, preprocessing methods such as PCA, ICA, or mutual-information-based band selection are routinely applied to and retain credible informative spectral features and reduce dimensionality [76].

6.6. Data Normalization and Standardization

Normalization and standardization are carried out to ensure consistency and address variability in hyperspectral data. Several methods are used for normalization. They include min–max scaling and vector normalization. The min–max scaling technique rescales the data linearly based on minimum and maximum values. On the other hand, vector normalization normalizes the spectral signature of each pixel by dividing it by its Euclidean norm.

In hyperspectral data, normalization minimizes the impact of external factors like lighting conditions and sensor-related variation. The effectiveness of normalization varies based on the type of dataset and analysis. Standardization, on the other hand, adjusts the data to have a mean of zero and a standard deviation of one. This ensures that all features contribute equally to the analysis [78].

7. Applications of HSI Systems in the Healthcare Sector

7.1. Detection of Tumors

HSI has been significant in detecting tumors by recognizing biochemical changes in tissue. Giannantonio et al. [79] in their study, extended on the utilization of deep-learning-based algorithms to visualize snap-scan from HSI technology and detect low-grade glioma margins. The images were taken from five patients with LGG intra-operatively and HSI with a surgical microscope (468–787 nm), and were annotated for further training. The results reflected an accuracy >90% after excluding over- and under-exposed areas, and using tiles that had low spectral variability [79].

Similarly, Weber et al. (2025) showed that intraoperative HSI, when correlated with FLAIR-MRI sequences, can accurately identify non-enhancing glioma tissue, offering a contrast-free tool for neurosurgeons to delineate tumor margins [54].

Hashimoto et al. [80] took liver pathological samples and experimented with a computer-based diagnosis. The samples were stained through hematoxylin and eosin (H&E). The result showed lower accuracy given the fact that RGB images of cytoplasm appear identical to the fiber. However, the application of HSI to those samples improved the results by 5% for cytoplasm and 24% for fibers compared to RGB images, and gave an accuracy of 100% in case of both fibers and cytoplasm (Figure 1) [80].

7.2. Image-Based Surgical Application

Cancers in the head and neck regions necessitate surgeries to ensure optimal healing. Current methods like microscopy and white light endoscopy have their own limitations which restrict complete resection. HSI has emerged as an instrumental avenue due to its ability to capture detailed spectral insights from 32 patients with head and neck squamous cell carcinoma. After surgical resection, patients underwent HSI imaging of the cancerous region that were annotated to utilize and further train a Graph Neural Network (GNN) and a Convolutional Neural Network (CNN). The results indicated that the HSI setup needed 12 min per sample image and demonstrated accuracy of 86% in predicting tumor due to AI and HSI. The figure attached below from the research highlight the accuracy of results through HSI Imaging in the first 20 patients (Figure 2) [81].

Another study evaluated HSI imaging and its diagnostic performance by collecting datasets from 200 lumpectomy specimens through hyperspectral cameras. The study classified algorithms to distinguish healthy tissues from tumors within a 0–2 mm margin. The findings reflected an accuracy of 83% and specificity of 78% to assess resection margins in ex vivo lumpectomy specimens which were imaged in 10 min. The results indicated the non-invasive and rapid alternative in practical intraoperative margin assessment (Figure 3) [82].

7.3. Burn Assessment and Wound Healing

HSI has emerged as a non-invasive procedure to conduct an assessment of burned areas and monitor wound healing. It offers deep insights into the depth of burn and perfusion to predict healing prospects. HSI provides contact-free and optimal procedure in clinical practice to estimate depth of burn wounds. The study describes that the hyperspectral spectra 3D perfusion parameters can be utilized on 72 h after thermal injury is described to determine the depth and the microcirculatory of burn. The HSI-based camera based on data processing methods calculates the 3D perfusion parameters of burned areas from the patient body. The images are used to estimate the oxygenation of hemoglobin volume fraction under layers of the injured skin and these indicators are used to recognize different stages and degrees of wounds. The results demonstrate that there are differences in the characteristics and features of wounds in perfusion parameters that depend on the degree of damage caused. The perfusion characteristic-based parameters are valuable in selecting the optimal treatment and reliable classification of burn degree [83].

HSI offers potential benefits in the dermatological domain by detecting and classifying various skin diseases such as psoriasis and melanoma. The study conducted an experiment by utilizing computer-aided diagnosis based on an AI algorithm, by collecting 1659 skin images of normal skin, cases of MF, AD, and PsO as a dataset. Advanced techniques such as XGBoost algorithms, U-Net Attention models enabled automatic segmentation of skin lesions. The images transformed to a spectral domain from color space. The model showed good performance in differentiating lesions with predictive capacity, and the result accuracy was ensured through the k-fold cross-validation value of 7. The results showed an F1-score of 90.08%, a specificity of 96.76%, and a sensitivity of 90.72% (Figure 4) [84].

Huang et al. [85] used hyperspectral imaging to gather wavelength insights on the skin cancer location by using datasets from the ISIC library. The dataset was classified as a test set and training set with YOLO-5 applied as a training model. The study implemented two models which were RGB classification and hyperspectral narrowband image (HSI-NBI) to compare their performance in accurately determining and classifying skin cancers. The performance of the model was evaluated through the parameter of five indicators, including the F1-score, accuracy, specificity, recall, and precision [85].

7.4. Endoscopy and Gastrointestinal Imaging

HSI-based endoscopy can provide potential benefits compared to conventional methods such as the clear visualization of biochemical and morphological insights that can improve disease diagnosis. HSI-based endoscopy allows for the detection of spectral profiles from pathological and normal tissues and quantification of blood oxygenation levels in tissue imitating phantoms [85]. The image distortions emerging from traditional endoscopic procedures has entailed a more advanced approach to visualize gastrointestinal images. The study addresses this challenge and demonstrates a hyperspectral endoscope that utilizes white light and co-registered hyperspectral image records [86]. The characteristic features enable enhanced image visualization and computationally compensate for distortions. HSI enables the development of an accurate hyperspectral data cube when an endoscope captures a lumen view. The evaluation of this approach promises high color resolution and enhanced temporal, spectral, and spatial resolution. The Figure 5 below shared in the study shows the detailed images of a pig’s esophagus [86].

Recent developments in transformer-based fusion models, such DCA-DAFFNet, show how adaptive feature fusion and deformable attention may successfully combine hyperspectral data with RGB or traditional endoscopic imaging. This method improves lesion grading and corrects for spatial misalignments during acquisition, providing increased resilience in the investigation of gastrointestinal pathology [87].

7.5. Tissue Oxygenation and Perfusion Monitoring

HSI offers non-invasive monitoring to check hemoglobin saturation, blood flow, and tissue oxygenation levels. This is achieved by analyzing the extent of light absorbed and reflected by tissue. The study conducts a review of NIRS with HSI approaches to compare their potential to detect vascular compromise in reconstructive flap surgery. Ninety flaps were conducted on ninety patients by using his, and 3662 flap surgeries were performed on 1970 patients through the NIRS approach. Through the literature review of the Embase and PubMed scientific databases, five HSI and sixteen NIRS studies were selected. The survival rate of NIRS and HSI flap was 99.2% (95% CI: 97.8–99.7) and 92.5% (95% CI: 83.3–96.8). Both approaches proved to be accurate, reliable methods to demonstrate oxygenation and perfusion monitoring [36].

8. Advanced Analytical Techniques for Hyperspectral Image Processing

8.1. Spectral Feature Extraction in Clinical Hyperspectral Imaging

A critical stage of HSI analysis is the spectral feature extraction. It is commonly applied in clinical settings for the purpose of obtaining precise differentiation among tissues. The HSI analysis captures a broad and deep spectrum of pixel level reflectance or absorbance data. The insights are used to infer physiological changes that are not discernible from conventional imaging sources. This section will deliberate on the theoretical foundations, spectral biomarkers, and practical implementation of spectral feature extraction [88].

8.1.1. Reflectance and Absorbance Computation

Most HSI systems are used in clinical setups work by converting raw intensity data into normalized reflectance [17]. This step helps to reduce the sensor noise and lighting variations. The equation is used to fine-tune white and dark reference images:

$$\mathrm{R}={\displaystyle \frac{{\mathrm{I}}_{\mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}}-{\mathrm{I}}_{\mathrm{d}\mathrm{a}\mathrm{r}\mathrm{k}}}{{\mathrm{I}}_{\mathrm{w}\mathrm{h}\mathrm{i}\mathrm{t}\mathrm{e}}-{\mathrm{I}}_{\mathrm{d}\mathrm{a}\mathrm{r}\mathrm{k}}}}$$

I_sample is the intensity recorded from biological tissues;

I_white represents the intensity from Spectralon;

and I_dark represents noise portion in camera.

There are some instances where clinical applications demand optical density such as layered or blood rich tissues; therefore, reflectance is converted into absorbance:

A(λ) = −log10(R(λ))

For diagnostic and classification modeling, these computations are integral to spectral analysis in clinical settings [72].

8.1.2. Spectral Biomarkers for Specific Tissues

Across the electromagnetic spectrum, tissue chromophores display a unique pattern of absorption that enables specialized wavelengths to be used as biomarkers in clinical settings [18].

Table 2. Clinically relevant biomarkers based on spectral features in HSI analysis.

Spectral Range (nm)	Dominant Absorber	Diagnostic Relevance
520–580	Oxyhemoglobin (HbO2)	Tissue perfusion, inflammation [89].
600–750	Deoxyhemoglobin (Hb)	Tumor hypoxia, ischemic regions [90].
900–1000	Water	Hydration status, edema, burn depth analysis [89].
1000–1400	Lipids and proteins	Tumor boundary delineation, soft tissue contrast [89].

summarizes commonly reflects physiological indicators based on spectral bands

These spectral range-based biomarkers have been substantiated by practical and controlled experimental conditions in wound imaging and brain tumor diagnosis applications. In histopathological analysis, feature-disentangled transformers, or FDTs, have recently been proposed as a way to extract spatial morphological characteristics from spectral biomarkers, such as lipid signatures or hemoglobin. By bringing transformer attention into line with biologically significant information, this disentanglement improves interpretability, which in turn increases clinical relevance and trust.

8.1.3. Construction of Spectral Vector

After the normalization of reflectance, spectral feature vectors are created by reorganizing hyperspectral data cube where a vector is connected to each pixel containing intensity values across the whole range of spectrum:

xi = [R1, R(2), …R(n)]

where n is the number of spectral bands. To classify algorithms and enable pixel-based association of various types of tissue, these vectors serve as crucial input during diagnostic monitoring of benign and malignant regions [89].

8.1.4. Application in Clinical Environments

Giannantonio et al. [79] utilized an HSI system-based snapshot with spectra range of 468–787 nm that was integrated with surgical microscope. HSI imaging helped in defining the boundaries of brain tumors. Spectral features were used to train the machine learning that were gathered from patients with first stage glioma cases. The study demonstrated high during practical surgery procedures. Further advancements introduced a public dataset comprising 36 hyperspectral images collected during human brain tumor surgeries. The dataset contains over 300,000 annotated spectral signatures for normal brain tissue, tumor regions, and vasculature, facilitating robust supervised training pipelines for medical AI systems [79].

8.1.5. Visualization of Spectral Discrimination

A clinical study on reflectance spectra reveals unique insights and differences between normal and pathological tissues such as altered levels of oxygenation in tumor tissues that resulted in reflectance profile changes between spectral range of 540–700 nm and demonstrated itself as a primary indicator in clinical diagnostic methods. According to the clinical experiment in the study, the hyperspectral imaging (HSI) system is shown being used during actual brain surgery [67]. It captures high-resolution spectral data directly from the brain surface in real time (A). Multiple hyperspectral images are acquired at different wavelengths from a glioblastoma-affected patient. These images reveal spectral variations between healthy and tumor tissues (B). Later, An HSI data cube is generated by stacking 2D images taken at different wavelengths. This 3D cube (x, y, λ) encodes both spatial and spectral information of the brain surface (C). An RGB image is reconstructed from the HSI cube for visualization. Tissue markers are placed: one on tumor tissue (left), the other on normal tissue (right) (D). Histopathology of the tumor tissue (glioblastoma) confirms its malignant nature. This serves as part of the ground truth for model validation (E). Histological image of normal brain tissue, providing a reference for healthy spectral patterns is used to compare against the tumor region (F). A gold standard tissue classification map is created. Each pixel is labeled into one of four categories: tumor (red), normal (green), blood vessels (blue), or background (black) (G). Spectral signature curves are plotted for each labeled tissue type (tumor, normal, blood vessel). They show the average and standard deviation of reflectance across wavelengths, aiding tissue differentiation (H) (Figure 6).

8.2. HSI and Dimensionality Reduction

HSI in clinical settings involves a system in which each pixel is made up of high-dimensional spectra vectors (with 50–200 bands). The rich insights are beneficial but also pose potential limitations such as noise, redundancy, and computational challenges that are mainly defined as dimensionality [68]. Therefore, to improve model performance and enable diagnostic monitoring, these issues are managed through techniques such as principal component analysis and independent component analysis.

8.2.1. Principal Component Analysis

It involves an unsupervised method to reduce dimensionality by converting the insights into principal components to indicate variations. In clinical environments, it is a widely used procedure to detect tumors and increase processing of HSI systems.

Formally, for a mean-centered data matrix $XRm\times n$ representing m pixels across n wavelengths, PCA computes eigenvectors $e_k$ and eigenvalues $k$ of the covariance matrix:

$$\Sigma ={\displaystyle \frac{1}{m}}XTX,$$

$$\sum e_k={\lambda }_k{e}_k$$

Spectral data are then projected onto the first $k$ principal components, where $k$ represents the number of components retained to capture the majority of the variance:

$$y_i=E_k^T{x}_i$$

The process helps provide rich insights in spectral features by reducing noise.

Ezhov et al. [90] extended the application of PCA by practically deploying it on intraoperative brain tumors using the HSI data matrix. The datasets were connected from HELICoiD. The study demonstrated that the initial PCA components accurately detected differences in chromophore particularly hemoglobin concentration and cytochrome in tumor tissue, aiding in identification of tumor tissues [91]. Moreover, PCA has been enhanced with several variations such as Super PCA, where PCA is experimented on contagious and very small regions using superpixels. Jiang et al. [5] used multiscale SuperPCA to enhance tumor detection and segmentation methods. Kernel PCA is another procedure under the same category that is applied on complex tissue reflectance features to capture insights on non-linear characteristics [5]. Ndu et al. [92] deliberated on correlation PCA fusion with clustering for the purpose of vein detection. The process involved combining PCA, custring and band selection based on correlation to improve vessel mapping in clinical settings [92].

8.2.2. Independent Component Analysis

This procedure focuses on separating multifarious signal into independent and underlying signal sources through statistical independence. The process aims to target and isolate mixed tissue spectra from chromophore signatures. Halicek et al. [93] used HSI data through independent component analysis to segment oral cancer tissue and demonstrated that the procedure accurately identified spectral features that were keratin-related. The results supported its application in clinical-grade tissue fingerprinting given the ability of the procedure to exhibit independent features with a higher correlation with histopathological labels.

$$S_i=W{x}_i$$

where components like S_i represents independent chromophore signals [93].

The equation denotes the separation process, where xi represents the observed mixed tissue spectra and $W$ is the unmixing matrix. Thus, it corresponds to the independent chromophore signals obtained after transformation.

8.3. Classification, Deep Learning Models, and Transformer Methods in Clinical HSI

After dimensionality is reduced, the deep learning algorithms, machine learning models, and transformer methods are used to classify spectral feature vectors to inform clinical based decisions such as inferring on tumor tissues. This stage is pivotal for translating spectral data into clinically actionable decisions, such as examining wound status and identifying tumor tissues.

8.3.1. Machine Learning Methods

Clinical HSI mainly deploys Spectral Angle Mapper and Support Vector Machines due to their computational efficiency and interpretability. These models works on angular commonalities between reference tissue signatures and spectral vectors:

$$\theta =cos-1 \left({\displaystyle \frac{t.r}{\Vert t\Vert \Vert r\Vert }}\right)$$

where t indicates spectral signatures of test and r represents reference pixels. Spectral angle mapper in combination with PCA for spectral band selection demonstrated 90% accurate results in detecting gastric cancer [62]. On the other hand, support vector machines aided in the classification of complex tissues particularly those in head-and-neck tumors, with specificity 89%, sensitivity 48%, and 76% accuracy. However, linear discriminant analysis (LDA) was most effective among all of these procedure as it demonstrated 93% specificity, 91% sensitivity, and 92% accuracy [94].

8.3.2. Deep Learning Approaches

Deep learning based approaches that involve the application of convolutional neural networks (CNNs) and their different types have been widely utilized in clinical HSI analysis given their potential in exploiting spectral and spatial insights simultaneously. For example, the 3D CNN model demonstrated 97% accuracy in distinguishing non-tumor tissue from tumor region in ex vivo and in vivo classification of tumors. It outperformed SVM with 95% accuracy and 2D CNNs with 96% accuracy [95]. The spatio-spectral variants of 3D CNN integrated spectral and spatial insights. The practical application involved a DenseNet-based architecture to detect laryngeal tumors. It demonstrated 81% accuracy in vivo [96].

A study inferred that CNNs demonstrated 86% specificity, 77% sensitivity, and 82% accuracy in seven head-and-neck cancer research [97]. Cruz-Guerrero et al. [97] applied hybrid models and various algorithms to combine traditional classifiers and deep learning approaches. The study classified histopathology HSI of the brain in relation with higher-level ensemble learning, conventional neural networks, spectral unmixing. The results demonstrated >90% specificity, >90% sensitivity, >90% accuracy in cross-validation during the testing phase [97].

By facilitating robust feature extraction, noise reduction, and spectral–spatial fusion, deep learning dramatically improves the diagnostic accuracy of hyperspectral imaging (HSI), addressing the complexity of biomedical data and bolstering broader healthcare diagnostics. This is demonstrated by the increasing clinical relevance of HSI in healthcare [53]. Another study tested combinations of spectral scaling, noise filtering, and patch sizes to optimize deep learning models for classifying colorectal cancer tissue in hyperspectral images [56].

8.3.3. Transformer Methods

Transformer methods have the potential to improve grading and classification in histopathology and similar medical imaging activities by providing a higher capacity to combine spatial and contextual information across an image. ViT-AMC being a transformer method combines Attention–Mechanism–Convolution (AMC) with Vision Transformer blocks, combining the two models in an adaptive manner while maximizing precision, readability, and feature duplication reduction. The benefit is its nature to improve grading accuracy in laryngeal histology by balancing the wider context of transformer models with the substantial inductive biases of convolution (local characteristics). The drawbacks of fusion include the model being heavier in training and inference, and the possibility of duplication or negative interference when the two branches acquire comparable characteristics. Pretraining to prevent overfitting on limited datasets and adaption to spectral tokens or band groupings would be necessary for applicability to HSI [98].

This second method, Swin-Transformer is utilized to better handle class imbalance and high inter-class similarity among lung adenocarcinoma histopathological subtypes.

Better identification of uncommon or challenging subtypes (higher F1, AUC, etc.) is one of its advantages, particularly in cases where datasets are unbalanced. Poor performance in particular groups where unusual types predominate the loss is a drawback. Adapting to spectral–spatial patches is necessary for HSI grading; localized loss may be beneficial for uncommon tumor classifications, but precise tuning is needed [99].

In order to achieve an interpretable grading of squamous cell carcinoma, FDT is another transformer method that separates out spectrum (“what material or biomarker”) and morphological (“shape, structure”) information. The benefit is that this kind of disentanglement can match model choices with pathologists’ visual cues or biological markers, thereby boosting interpretability and trust. Disentanglement, however, increases modeling complexity, may require strong priors or auxiliary supervision, and may suffer from overfitting or unstable separation in the event of insufficient data [98].

Another technique utilized in healthcare is LA-ViT, which aims to emphasize target semantic regions without overlearning irrelevant background by lessening the importance of background (“low-effect”) regions.

Better classification in datasets with reduced background noise and improved interpretability are some of its benefits. The drawback includes less flexibility since a parameter-free restriction may suppress vital context that appears to be in the “background.” Another drawback is that attention accuracy may be compromised [100].

9. Clinical Utility and Emerging Impact Areas of HSI

Along with the proven application of HSI technology in surgical guidance and oncology, it has been demonstrating significant potential in other domains of disease diagnosis that have the ability to impact patient outcomes and clinical decisions. The emerging aspects reflect the diverse applicability of HSI systems not only as intraoperative tools but also in their ability to provide translational research opportunities for wider application in healthcare.

9.1. HSI Utility in Organ Transplantation

The applicability of HSI technology has been analyzed in the domain of organ transplantation. As the technology can serve to assess the quality of graft quality and provide viability data in real-time. HSI has been demonstrated to provide quantified data on hepatic steatosis during a normothermic machine perfusion of liver grafts. While providing wider coverage of tissue, it indicates a strong correlation with histological assessments and also eliminates the probability of sampling bias [40]. HSI has also been deployed in the evaluation of perfusion in pancreas allografts after transplant and has helped to identify deficit in tissue oxygenation. The research has enabled its utility in confidently taking decisions on graft acceptance and surgical planning [52].

9.2. Microcirculatory Monitoring in Critically Ill Patients

Hyperspectral imaging has been deployed to non-invasively monitor and asses bedside microcirculation during critical illness. Bedside hyperspectral imaging systems have shown lower perfusion index and tissue oxygen. The technology has also demonstrated increased tissue water index. These spectral changes have a strong correlation with the established ICU severity index. The immediate data from ICU patients has helped in predicting mortality (AUROC 0.72) as well as sepsis (AUROC 0.80). When integrated with other clinical variables, it showed further improvements to provide guidance on vasopressor interventions and fluid therapy [51].

9.3. Screening of Ophthalmic Disease

HSI has been considered more accurate than conventional fundus photography to provide retinal imaging non-invasively and measure oxygen saturation in arteries and veins more deeply. In case of diabetic retinopathy where changes occur in distinct spectral patterns that align with retinal oxygen at different stages of disease such as proliferative DR versus the pre-proliferative stage. The classification accuracy was more than 90%. The results demonstrates the potential of HSI in detecting and monitoring retinal microvascular disorders in the early stages [50].

9.4. Neonatal Jaundice Detection

Traditional approaches to jaundice detection have been built on invasive procedures like blood tests. However, the development of non-invasive biosensors has provided a potential avenue for HSI deployability. In order to detect bilirubin levels through optical absorption, spectral imaging-based biosensors have been used that offer a non-invasive diagnosis of neonates. Specifically, absorbance-based biosensors, and optical biosensors, such as fluorescence have demonstrated the most accurate results in terms of sensitivity and specificity. However, gaps exists in the detection of biomarkers other than bilirubin, validation of biosensors in clinical setups, and standardization of biosensor measurements [101].

9.5. Evaluation of Perfusion Quality in Haemodynamic Therapy

Hemodynamic therapy has been resourceful in improving organ perfusion and microcirculatory tissue. Hyperspectral imaging has shown immense potential in enabling noninvasive microcirculatory monitoring. During hemodynamic shock, HSI demonstrated dynamic changes in perfusion quality and tissue oxygenation levels and also aided in the indication of resuscitation effectivity. The research indicated strong correlation between HSI skin and kidney parameters. These measurements can offer estimates regarding organ oxygenation impairment by monitoring skin. The applicability of HSI microcirculatory monitoring in hemodynamic management can open up new avenues for more advancements in individualized resuscitation strategies [102].

10. Challenges and Limitations

10.1. Data Complexity and Volume

HSI faces challenges related to data complexity and volume. Its high dimensionality of HSI data leads to increased estimation errors [50]. Redundancy in contiguous spectral bands makes feature extraction increasingly difficult. The presence of noisy bands can result in information loss [18]. For instance, the processing of large volumes of spectra database related to molecular biomarkers and tissue types, including epithelial/mucous tissue, ocular tissue, aorta, liver, lung, breast tissue, cartilage, skin and subcutaneous tissue, myocardium, muscle, and head/brain tissue enables the detection and differentiation between possible oxygenation level of tissues and deoxygenated blood. It also demonstrates various types of including fatty tissues, bile duct that surrounds it [18].

10.2. High Computational Requirements

HSI requires high computational power due to its high-dimensional data and the complexity of processing large amounts of spectral and spatial data. This requires advanced hardware to manage the huge data volumes. This and high spatial variability in HSI datasets increase its computational load [83]. The challenge of quickly computing large datasets is evident in the task to generate a map of tissue types, extract high-quality diagnostic insights and uncover disease-related endogenous substances [103].

10.3. Cost and Accessibility of Equipment

The high cost of powerful spectrometers and quality cameras in HSI systems poses financial barriers to their adoption. Additionally, the systems require huge computational power for both data storage and processing, which further increase the overall cost [104]. Significant research on development of low-cost HSI systems focus on miniaturization or modification of spectral module by combining it with spatial module and utilization of diffusers or meta surfaces, along with LED arrays, AOTF and LCTF devices for multispectral illumination. Sensors are also experimented to increase sensitivity range of spectra or include a spectral selection module. These experiments are carried out to reduce the cost of components inside HSI systems and expand their applicability across healthcare sector. Despite these efforts, there is no success given the cost of technology and materials [105].

10.4. Calibration and Standardization Issues

Calibration models rely on limited datasets, which may fail to generalize effectively to real-world scenarios. The complexity of light-tissue interactions as well as the variability in spectral–spatial features require advanced validation methods to ensure accuracy. There is a need to develop a more standardized database for HSI systems that combines metadata related to acquisition and calibration that will enable independent validation of spectral data due to enable smooth processing of algorithms. It will establish validation technique in vivo along with in silico and ex vivo clinical cases [44].

Achieving real-time processing can be a major challenge since processing such large datasets requires high computational power. This power requirement may exceed the capacity of existing systems [81]. Additionally, despite its extensive utilization in research, most studies have limited proof of the application and integration of HSI in clinical environments where these systems are used in routine. There is a need for expansion of clinical trials combining HSI Systems to produce constructive surgical outcomes in healthcare domain. As reflecting on the potential of this technology for health sector is impossible with large scale integration in healthcare setups. Enabling hardware refinements will enable switching between spectral data acquisition and real-time color. Such endeavors will for-go the necessity of developing an expert set-up to maximize clinical uptake [106].

11. Recent Advances and Emerging Trends

HSI has gone through major advancements triggering various challenges like the vast data volumes it generates. This is caused by capturing detailed spectral information, which necessitates advanced data processes, management techniques, computational resources, and storage [82]. Recent studies show the high potential of integrating HSI with ML and AI algorithms to address these challenges.

11.1. Integration with Other Technologies

Significant developments in HSI have led to a number of problems, such as the enormous amounts of data it produces. In order to overcome such obstacles, recent research indicates that HSI can be effectively integrated with ML and AI algorithms.

For instance, HSI combined with deep leaning models such as GANs and CNNs have increasingly aided in the mapping of tissues and classification of tumor regions. The integration of such models has reduced clinical workloads due to AI-based automated interpretation. Complex clinical conditions like neurodegeneration have been more accurately diagnosed due to multimodal fusion techniques that include HSI + MRI [45]. Researchers developed shallow autoencoders to reduce dimensionality and extract robust spectral-spatial features, improving cancer tissue classification from medical hyperspectral datasets [107].

11.2. Miniaturization and Portability

The miniaturization and portability of HSI systems have advance in recent years and are being used various environmental and field-based applications. These advancements are crucial in environmental monitoring where the development of compact and low-cost HSIs that are easy to use is critical [12]. For instance, smartphone-based endoscopy gadgets have been used to combine HSI technology with smartphones. It involved hemodynamic monitoring and spectral mapping based on color code which is an endeavor toward cost-effective HSI application in emergency situations [46].

There have also been recent efforts to integrate consumer-grade technologies into HSI systems. These changes are meant to lower their operational costs without degrading data quality [108]. The trend toward the miniaturization of HSI systems also include user-friendly platforms, which are suitable for non-specialist operators [30]. Advances in sensor technology like push-broom and whisk-broom sensors now make it possible to achieve high spatial and spectral resolution [82]. As miniaturization advances, the HSI systems are also becoming more capable and less costly [109]. Miniaturization has enabled the development of a first miniature rigid endoscope that combines the HSI technique. It has been considered suitable for laparoscopic operations and ENT to provide insights on perfusion metrics and tissue oxygenation levels [48].

However, several barriers must be realized such as that smartphones are not designed for narrowband imaging. Secondly, most prototypes achieve only 16–32 spectral bands which is significantly lower than laboratory HSI systems (>100 bands). Similarly, smartphones struggle with the data intensity of HSI, raising challenges in real-time analysis and storage [110].

11.3. Enhanced Data Processing Algorithms

Recent advancements in HSI have led to increased innovations in machine learning and artificial intelligence. Machine learning and deep learning are highly useful in analyzing hyperspectral data. They have enhanced automation in feature extraction, which reduces computation time and increases accuracy. Reinforcement learning happens to be another emerging algorithm. It is a human-like learning algorithm that learns through trial and error, and has been considered a promising approach in HSI systems [33,111]. For example, HSI-based technology has combined spectral and spatial features through machine learning algorithms to provide more accurate results than SVMs and an improved detection of tumor margins along with promoting non-invasive diagnosis of cancer regions [95].

11.4. Cloud Computing and Big Data Solutions

Cloud computing and big data solutions have been used to address the HSI data storage and processing challenges. Cloud-based platforms have been incorporated to ease the storage and processing of these large-scale datasets. The integration of cloud computing with big data analytics also offers flexibility when handling real-time data processing [46]. For example, cloud-based hyperspectral imaging analysis has been used in remote clinical setups particularly in dermatological sections to support healthcare expansion in underdeveloped regions along with reducing computational burden on local machines due to edge computing [47]. Additionally, clustering-based compression has merged as a cloud storage technique has emerged in HSI application in clinicals setups to create HSI-based archives on clinical data and support medical AI training through cross-institutional data sharing [112].

11.5. Advances in Sensor Materials and Design

Advances in sensor materials and design have improved the performance of HSI systems. Innovations in sensor materials have enhanced the ability to capture detailed spectral data. Such advancements enable improved precision in the measurements of radiation absorption and emission. This has helped in generating hyperspectral images in greater quality. Typically, the sensitivity and ability to capture finer spectral details increase as HSI systems evolve, leading to more efficient data collection and analysis [113,114]. Subcutaneous vessel visualization through hyperspectral imaging provides insights on skin perfusion and assessing vascular depth which are beneficial in pre-operative scenarios. The system enables calibration across a diverse range of skin tones to improve diagnostic equity in clinical setups [115].

12. Potential Research Areas

12.1. Advanced Machine Learning Techniques

HSI shows high potential in identifying early plant stress and disease. This depicts its capacity to analyze biotic and abiotic factors based on high-fidelity spectral data. Combining HSI with machine learning models can help enhance classification accuracy, automate stress detection, and improve the predictions of disease [114].

12.2. Real-Time Data Processing and Analysis

HSI presents major opportunities to advance the technology in real-time data processing and analysis. The focus could be on enhancing both speed and efficiency [105]. In this regard, future research could focus on optimizing hyperparameter tuning in HSI models because this would improve prediction accuracy, increase convergence speed, and broaden generalization.

12.3. Sensor Development and Optimization

Potential research areas in HSI regarding sensor development and optimization lie in how to create more sensitive, accurate, and cost-effective sensors. Advances in hyperspectral sensor technology can help enhance imaging performance in various applications including biomedical diagnostics. This would be required since efforts to improve sensor configurations and hardware components seem to be relatively limited [116].

12.4. Integration with IoT and Smart Systems

There have been significant efforts to enhance HSI sensor development and optimization. Advancements in hyperspectral sensor technology can be used to enhance imaging performance in various applications including remote sensing and biomedical diagnostics. More research should be carried out toward improving sensor configurations and hardware components since research in these areas seems relatively limited [117]. There needs to be more research on the integration of HSI with the Internet of Things and smart systems.

12.5. Enhanced Calibration and Standardization Methods

The clinical and experimental reliability of HSI is based on effective calibration and standardization methods. According to Datta [39], HSI faces challenges from external factors like lighting conditions and atmospheric variability, as well as imaging distances [39]. These can affect the data quality and consistency. Calibration and standardization are thus needed to address these issues. Pre-acquisition calibration techniques like white and dark reference calibration, as well as advanced post-processing algorithms are needed to normalize spectral data.

12.6. Standardization of HSI Protocols in Clinical Setups

The clinical studies demonstrates that there are inconsistencies in the process of image acquisition parameters across wide range of HSI systems. The absence of standardized calibration protocols makes it difficult to infer on the credibility of each system used. There is a need to develop a uniform protocol related to HSI analysis to provide consistent diagnosis in clinical setups. The inconsistencies revealed by the varying wavelength of various HSI systems reflects on the need for widely recognized protocols and tools in clinical HSI that holds shared spectral libraries and databases to encourage standardization [118].

12.7. Wearable HSI Devices

Research has opened an avenue for the development of portable HSI devices through the integration mosaic arrays, 3D-printed optics and Fabry–Perot interferometers. Such technological breakthroughs will enhance real-time bedside monitoring as well as intraoperative application. Moreover, the development of portable HSI technology will require shrinking camera size to 100 g to provide images of wound as well as surgical guidance. Such innovations will support the creation of cost-effective and fast acquisition systems that can be used in mobile clinical setups. However, their practical application is hindered by economic and technical barriers such as the miniaturization of HSI sensors have created fabrication complexity and increased production costs which limits its affordability in routine healthcare setups. Additionally, real-time processing demands that are emerging from the necessity of handling vast range of spectral bands per frame presents itself as a constraint for portable devices like bedside monitoring or those used in operating rooms [53].

12.8. Real-Time Neuro-Vascular Imaging

The application of HSI technology can be explored and consolidated in one of the most significant clinical conditions which is brain hemodynamics. These technologies can serve to map stroke or tumor perfusion in vivo trough the utilization of advance silicon chips that can capture brain oxygen dynamics in real time. More advancements can open avenues of intraoperative neurovascular monitoring and quick functional images during surgical phase. However, Challenges persists in maintaining temporal resolution given that cerebral hemodynamics alters in milliseconds. Moreover, the miniaturization of sensors suitable for neurosurgical microscopes holds cost and manufacturing barriers and real-time data processing remains computationally demanding [119]. Another review maps how hyperspectral imaging is being applied in early disease detection, tissue biochemical analysis, and drug/food quality testing, while noting barriers to clinical adoption [120].

13. Future Directions

13.1. Potential Breakthroughs

The future directions for HSI should align with the need to address the challenges of miniaturization, cost reduction, as well as enhanced sensor performance. This would be a major breakthrough in increasing its adoption in various applications [106]. Therefore, the focus could be on the development of compact and lightweight configurations for better configurations.

13.2. Creation of Disease-Specific Databases

Future work should enhance the standardization of spectral datasets that will aid in consistent training of machine learning when it comes to specific diseases like retinopathy and cancer. Standardization will also facilitate reproducibility across institutions and emergence of comparative studies due to automated feature extraction and improved diagnostic specificity [121].

13.3. Integrating HSI in Clinical Setups

In order to improve clinical applicability of HSI technology, its deployment across diverse range of clinical decisions and support systems is necessary to provide diagnostic insights. This will generate and enable holistic data integration due to interoperability with electronic health records of patients due to reduced training barriers from tailored dashboards and graphical interfaces [42].

13.4. Clinical Validation and Regulatory Standards

In order to ensure regulatory standardization of HSI-enabled diagnosis, there is a gap for clinical validation that can be achieved through specialized acquisition protocols, multicenter research, and outcome-based metrics [45]. There is a need for research to focus on generating high-quality evidence on specificity, sensitivity, and cost-effectiveness of such technologies in clinical environments, particularly in the domains of dermatology, oncology, wound care, and ophthalmology [45].

14. Conclusions

14.1. Summary of Key Points

HSI has great potential in healthcare and other sectors including agriculture, environmental monitoring, and biomedical applications. In this paper, the key areas of focus included the integration of machine learning and deep learning techniques to improve data analysis and enhance classification accuracy. The evaluation covered aspects like the development of real-time processing algorithms to handle dynamic environment. Innovations in sensor technology has also been highlighted as areas of research to help improve efficiency and reduce the over costs of operating HSI systems. Owing to its potential, HSI can be applied to various emerging fields, including urban planning and food safety.

14.2. Implications for Research and Industry

The study’s advancements highlight HSI’s potential across research and industry, especially in medicine, enabling more precise surgeries and improved cancer treatment.

14.3. Future Research Directions

Future HSI research should concentrate on optimizing and simplifying sensors for small, affordable solutions in order to broaden its uses. In addition to established calibration procedures to guarantee consistency across systems, developments in AI and deep learning will likely improve spectral analysis.

14.4. Final Thoughts

Typically, HSI is a powerful tool with high potential for triggering advancements in various industries and scientific research. The development of more efficient sensors and improvements in data processing algorithms will soon increase the adoption of HSI in numerous fields. However, there is an increasing need to address the highlighted challenges related to sensor optimization, calibration, and regulatory standards.