A Systematic Analysis of Neural Networks, Fuzzy Logic and Genetic Algorithms in Tumor Classification

Abstract

This study explores existing research on neural networks, fuzzy logic-based models, and genetic algorithms applied to brain tumor classification. A systematic review of 53 studies was conducted following PRISMA guidelines, covering search strategy, selection criteria, quality assessment, and data extraction. Articles were collected from three scientific databases: Web of Science, Scopus, and IEEE. The review primarily focuses on practical contributions, with most studies emphasizing applications over conceptual insights. Key methods in the field demonstrate significant impact and innovation. Commonly used training and testing mechanisms include dataset splitting, augmentation, and validation techniques, highlighting their widespread adoption for performance evaluation. The analysis of evaluation metrics shows that accuracy and the DICE score are the most frequently used, alongside sensitivity, specificity, recall, and other domain-specific measures. The variety of metrics underscores the need for tailored approaches based on dataset characteristics and research objectives. By highlighting trends, challenges, and research gaps, this review provides actionable insights for advancing BTC research. It offers a comprehensive overview of techniques and evaluation methods to guide future developments in this critical domain.

1. Introduction

Brain tumors are among the most dangerous types of cancer and require precise and timely classification to improve patient outcomes. Traditional methods of brain tumor diagnosis rely heavily on manual interpretation by radiologists, which can be time-consuming, subjective, and prone to errors. Early diagnosis is crucial in determining the appropriate treatment path, including surgery, radiotherapy, or chemotherapy [1]. In recent years, integrating AI-based technologies—such as artificial neural networks (ANNs), fuzzy logic-based neural networks, and genetic algorithms—into brain tumor diagnosis has transformed the field. These AI-based technologies enable automatic, high-speed, and accurate tumor classification, enhance clinical decision-making, and reduce human error.

Artificial neural networks (ANNs) have received significant attention in medical image analysis and classification due to their ability to solve complex computational problems, particularly in medical diagnostics that require high accuracy [2]. One of the key applications of ANNs is brain tumor classification (BTC). BTC systems rely on medical imaging data, such as MRI scans, to detect, diagnose, and plan treatment for brain tumors at an early stage [3]. Additionally, BTC systems assist radiologists and clinicians in accurately distinguishing between different types of brain tumors, such as gliomas, meningiomas, and pituitary tumors [1]. The typical workflow of BTC algorithms consists of several key stages: (1) data acquisition, where MRI scans are collected; (2) preprocessing, which includes noise reduction and image normalization to enhance quality; (3) segmentation, which identifies regions of interest, particularly tumor areas; (4) feature extraction, where characteristics such as tumor size and texture are identified; (5) classification, using techniques like ANNs, fuzzy logic-based neural networks (FLNNs), or genetic algorithms (GAs) to determine the tumor type; and (6) evaluation, where the model’s performance is assessed based on accuracy, precision, recall, and other relevant metrics [4].

Deep learning models, a subset of ANNs, have demonstrated their ability to learn complex patterns in data. Numerous studies have explored the effectiveness of popular deep learning models, such as convolutional neural networks (CNNs), in BTC [5]. CNNs are particularly well-suited for image processing tasks and have been widely applied in medical imaging. However, despite their success, several challenges remain. One major challenge is the variability in datasets used across different studies, often leading to inconsistent performance when applied to new or diverse patient populations. Another challenge is comparing results across studies due to the lack of standardized evaluation metrics and transparent training and testing mechanisms [6].

Fuzzy logic-based neural networks (FLNNs) represent another promising BTC approach. FLNNs combine the strengths of fuzzy logic and neural networks to address uncertainties and imprecise data in medical image classification [7]. Fuzzy logic enables the system to interpret data more flexibly using linguistic variables such as “small”, “medium”, or “large” [8]. FLNNs enhance the accuracy of BTC systems by mitigating noise and ambiguity often present in medical images. Combining fuzzy rules derived from expert knowledge and neural network learning capabilities improves the system’s interpretability and robustness.

Another approach involves using GAs inspired by natural selection. GAs are employed to optimize the parameters of ANNs and to select the most relevant features from MRI images for accurate classification. Integrating GAs into ANNs enhances BTC accuracy by optimizing feature selection and adjusting the hyperparameters of ANN models. Moreover, the ability of GAs to evolve through crossover and mutation processes enables efficient optimization of complex models such as ANNs [9].

The primary challenge in applying ANNs, FLNNs, and GAs in BTC lies in the diversity and inconsistency of methods, datasets, and evaluation metrics. Additionally, the lack of standardized datasets and preprocessing techniques complicates the assessment of model generalizability. This systematic review follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [10]. The study aims to consolidate the current body of knowledge in this field by extracting and analyzing key data from 53 relevant studies. It examines various types of neural networks, fuzzy logic-based models, genetic algorithms, the datasets used, segmentation approaches, and evaluation metrics. This review provides insights to guide future BTC research and development by identifying common trends, research gaps, and challenges.

This study aims to answer the following questions:

• What are the trends in applying ANNs for segmentation tasks, as reflected in the types of ANNs, datasets, and segmentation approaches used in the recent literature?
• How do different training and testing mechanisms and evaluation metrics correlate with performance outcomes (e.g., accuracy, precision) in ANN-based segmentation models across the selected papers?

Unlike previous systematic reviews that primarily focus on specific models, this study uniquely integrates and compares artificial neural networks, fuzzy logic-based neural networks, and genetic algorithms within the same systematic framework. Moreover, it contributes by classifying studies into conceptual and practical categories, detailing the train-test mechanisms, segmentation approaches, and evaluation metrics.

The remainder of this paper is organized as follows: Section 2 explains the study’s methodology. Section 3 presents the descriptive analysis of the selected studies. Section 4 provides qualitative insights and discusses key findings. Finally, Section 5 concludes the research with recommendations for future work.

2. Methodology

The PRISMA methodology is one of the most widely used frameworks for reporting meta-analyses and systematic reviews [10]. It has been applied in various research fields, including deep learning, machine learning methods, and IPv6 security issues [11]. Figure 1 illustrates the methodology followed in this paper.

2.1. PRISMA Basic Spets

The PRISMA framework consists of four main stages:

• Search strategy;
• Selection criteria;
• Ensuring result quality.
• Data Extraction

Systematic research requires defining and developing a search strategy to identify relevant literature. This strategy includes three major scientific databases: Web of Science, Scopus, and IEEE. These databases are among the largest repositories of scientific and technical publications and contain a wealth of high-quality research. One of the search terms used in the databases was “brain tumor”. In the initial research phase, the search scope was limited to studies published between 2014 and 2024.

The study included research from all countries but was limited to publications written in English. In the Web of Science database, the “publication topic” option was used as a filter to focus on image processing with medical images. This refinement reduced the number of studies from 113 to 99, as 14 were excluded. In the IEEE database, the “publishing topic” criterion was applied to focus on medical image analysis, excluding 18 studies from a total of 101, leaving 83 relevant studies. For the Scopus database, the selection criterion was based on “keywords”, yielding 41 relevant papers out of 58, with 17 papers excluded due to irrelevance. The final number of selected research papers across all databases was 223.

Table 1. Selection criteria.

	Web of Science	IEEE	Scopus
Keyword:	“brain tumor”	“brain tumor”	“brain tumor”
Criterion applied:	“publication topic”	“publishing topic”	“keyword”
Value of criterion:	Image processing with a medical image	Analysis of medical image	Image processing in the medical field
Records returned:	113	101	58
Record excluded:	14	18	17
Records extracted:	99	83	41

presents the selection criteria applied in this study.

Three main steps were taken to improve the quality of the search.

Firstly, the “document type” field was used in the Web of Science and Scopus databases to filter search results, while the “document identifier” field was applied in the IEEE database. Additionally, a filter was implemented to exclude review articles and conference papers. As a result, six articles were excluded from Web of Science, two from Scopus, and four from IEEE.

The second step was removing the duplicates, which was performed using the duplicate removal features in Microsoft Excel in two stages:

A. Removing duplicates within individual databases

• Web of Science: in total, 5 papers removed;
• IEEE: in total, 4 papers removed;
• Scopus: in total, 10 papers removed.

B. Matching across all three databases and removing duplicates

• Web of Science: in total, 15 additional papers removed;
• IEEE: in total, 13 additional papers removed.

This step excluded 26 papers from Web of Science, 21 from IEEE, and 12 from Scopus.

The last step, evaluation step was divided into two parts. The first part focused on analyzing the abstracts of the papers, while the second part examined citations and the year of publication.

A.

Extracting articles that employed magnetic resonance imaging (MRI)

The abstract summarizes the research work, making it useful for identifying relevant studies. Analyzing and evaluating the abstract content helped isolate research that specifically employed MRI. Any study that did not rely on MRI for conducting experiments was excluded. As a result, 28 articles were excluded from Web of Science, 22 from IEEE, and 10 from the Scopus database.

B.

Extracting highly cited articles

This step involved identifying and selecting articles with high citation counts from the past decade (2014–2024). As a result, 12 papers were excluded from Web of Science, 29 from IEEE, and 10 from the Scopus database.

The statistical summary of the studies is presented in

Table 2. Stage of maintaining quality.

Task	Stepping	Excluded Papers
		Web of Science	IEEE	Scopus
Filter	Counting articles, reviews, and conference papers	6	4	2
Removal duplicated	Exclusion based on duplicates in the same database	5	4	10
	Exclusion based on duplicates between the three databases	15	13	0
Evaluation	Analyzing the abstract papers	28	22	10
	Citations and the year of publication of papers	12	29	10

, under the Stage of Maintaining Quality section.

2.2. Data Analysis

The data analysis focused on identifying key insights concerning the review’s principles and research questions. Classification and analysis were conducted based on titles, keywords, and abstracts, as they contained the primary data and were examined rigorously. This study followed the same approach as Abdullah et al. [11].

Each article was classified as either a “conceptual contribution” or a “practical contribution”.

• A conceptual contribution refers to descriptive, comparative, analytical, or review-based research;
• A practical contribution indicates that the research involved designing, developing, implementing a program, or presenting a novel algorithm.

Another level of analysis was based on the research domain. This study explores brain tumor classification, focusing on the parameters used in related research. For instance, some researchers aim to classify and detect brain tumors using MRI images and apply machine learning or deep learning techniques. Thus, based on the applied methods, this classification approach primarily relies on MRI images for brain tumor detection.

The study considers two primary perspectives:

• The “field” perspective examines the specific parameters researchers worked with;
• The “parameter” perspective reflects the targeted application areas.

Additionally, a third classification focuses on the specific problems the studies aimed to address, such as data preprocessing or brain tumor classification.

3. Results

This section presents the results of applying the PRISMA framework, including a descriptive analysis and classification of the literature.

3.1. Descriptive Analysis

In response to the first research question, “What are the trends in applying ANNs for segmentation tasks, as reflected in the types of ANNs, datasets, and segmentation approaches used in the recent literature?” several descriptive statistical analyses were conducted. The following subsections detail these statistical analyses.

3.2. Trends in Citations and Publications over Time

This subsection examines citation trends and the number of publications from 2014 to 2024 across different indexing platforms, including IEEE, Scopus, and Web of Science. Figure 2, Figure 3 and Figure 4 illustrate these trends, providing insights into the growth and impact of research on ANNs for segmentation tasks over the past decade.

The abnormal rise in citation trends observed in Figure 2 can be attributed to the influence of several highly cited research papers published in 2018 and 2019. Notably, studies such as Wang et al. (2018) [12], which introduced advancements in interactive medical image segmentation using deep learning with image-specific fine-tuning, and Mallick et al. (2019) [13], who developed a deep wavelet autoencoder-based approach for brain MRI cancer detection, significantly contributed to the spike in citations.

Additional works by Hossain et al. (2019) [14] and Hemanth et al. (2019) [15] on brain tumor detection using convolutional neural networks and machine learning approaches further fueled this sharp increase, reflecting a surge in interest and acknowledgment of these impactful studies.

For Figure 3, the abnormal increase in citations can be attributed to the influence of highly cited research papers, particularly the comprehensive survey by Litjens et al. (2017) [16], titled “A Survey on Deep Learning in Medical Image Analysis”. This foundational work provided an extensive overview of deep learning advancements in medical imaging and became a critical reference for subsequent studies in the field. Its widespread citation underscores the survey’s pivotal role in shaping research directions and methodologies, leading to the marked rise in citation counts observed in Scopus.

The figures illustrate a consistent upward trend in citations and publications from 2014 to 2024 across IEEE, Scopus, and Web of Science. This steady increase highlights the growing relevance and research interest in artificial neural network (ANN) applications for segmentation tasks, with notable growth in recent years. The patterns depicted emphasize the broader academic community’s engagement and the significant impact of ANN advancements on segmentation research, reflecting the field’s evolving influence and importance in scientific literature.

3.3. AI Techniques Used in Brain Tumor Classification

This subsection provides a detailed analysis of the AI techniques applied in the reviewed literature for brain tumor classification tasks. The findings, illustrated in Figure 5, Figure 6, Figure 7, Figure 8 and Figure 9, reveal a predominant use of various AI methods.

Figure 5 presents the overall distribution of AI techniques, indicating that neural networks account for 69% of all approaches. Among these, convolutional neural networks (CNNs) are the most widely utilized, as depicted in Figure 6 (56%), followed by U-Net (37%) and transformers (2%).

Figure 7 details machine learning methods, highlighting supervised learning techniques, such as naïve Bayes, support vector machines (SVM), and k-nearest neighbors (KNN). SVMs contribute 54% to the supervised learning category, followed by KNN (15%) and extreme learning machine (ELM) (15%).

Unsupervised techniques, shown in Figure 8, include K-means and fuzzy c-means (FCM) clustering, each representing 50% of the methods.

Figure 9 presents the usage of other algorithms, such as generative adversarial networks (GANs) and adversarial networks like MTANS, which represent emerging trends but are less frequently applied.

This comprehensive analysis underscores the diversity and evolution of AI methodologies in advancing brain tumor classification.

3.4. Dataset Analysis

After examining AI techniques, this subsection explores the types of datasets used in the reviewed studies. As depicted in Figure 10, the analysis reveals that 13 datasets were sourced from hospitals, emphasizing real-world medical data for developing and validating AI models. Among these databases that were used in the articles are as follows: Tianjin Universal Center of Medical Imaging and Diagnostic (TUCMD), Seoul National University Hospital, and The Cancer Imaging Archive, covering more than 2000 patients from over 200 institutions. In contrast, 38 datasets were obtained from publicly available repositories, such as Kaggle (BRATS 2013, BRATS 2014, BRATS 2015, BRATS 2016, BRATS 2019, BTD20, BraTS dataset, BraTS 2017, BraTS 2018, BraTS 2021, FigShare, SARTAJ, and Br35H). The databases covered different ages, from newborns to adults, as well as different types of tumors, highlighting the accessibility and utility of internet-based data resources.

This dataset analysis underscores the importance of diverse and comprehensive data in advancing brain tumor research.

In addition to model-related challenges, computational issues within datasets also pose significant barriers. These include the need for extensive data cleaning to address noise, labeling errors, and format inconsistencies. Missing data are another common issue that can affect model training and evaluation. Furthermore, large dataset volumes require substantial computational resources for storage, processing, and training, which may not be accessible in all research settings. These challenges must be addressed to ensure reliable and scalable model development.

3.5. Segmentation Approaches

This subsection provides an overview of the segmentation approaches used in brain tumor classification, as presented in Figure 11. The analysis reveals that ANNs are the most widely adopted, accounting for 47% of all segmentation techniques. Within this category, commonly used approaches include CNNs, U-Net, context-guided attentive conditional random fields (CRFs), transformers, and the NeuroXAI framework.

Machine learning approaches constitute 18% of the segmentation methods. Several algorithms are employed, such as fuzzy c-means clustering, K-means clustering, and GANs.

The remaining 34% of approaches fall under “Other Techniques”. These include essential image processing methods such as image normalization, feature extraction, filtering and denoising, edge detection, and mathematical transformations. These techniques are crucial for preparing and refining medical images.

This distribution underscores the dominant role of ANNs while highlighting the complementary use of machine learning and image processing techniques in brain tumor segmentation research.

3.6. Literature Classification

A comprehensive qualitative analysis was conducted in response to the first research question. The following subsections describe the steps taken to complete this analysis.

3.6.1. Classification of Articles into “Conceptual Contribution” and “Practical Contribution” Groups

Each article was classified as either a “conceptual contribution” or a “practical contribution”.

• A conceptual contribution refers to research that involves description, comparison, analysis, or review-based studies on the topic;
• A practical contribution indicates that the research involved designing, developing, or implementing a program or presenting a novel algorithm.

This study includes 53 research papers, analyzed based on their research methodology. The articles were categorized as either conceptual or practical contributions, as shown in

Table 3. Article classification.

No.	Authors	Approach	Type
1	Díaz-Pernas et al., 2021 [17]	CNN, sliding mechanism	Practical
2	Ghafourian et al., 2023 [18]	Social spider optimization (SSO) algorithm	Practical
3	Menze et al., 2021 [19]	U-Net, intensity normalization, spatial harmonization	Conceptual
4	Kurdi et al., 2023 [20]	Fuzzy c-means clustering	Practical
5	Vankdothu & Hameed, 2022 [5]	K-means clustering, gray level co-occurrence matrix (GLCM)	Practical
6	Daimary et al., 2020 [21]	CNN, U-Net	Practical
7	Reddy et al., 2018 [22]	K-means clustering	Practical
8	Ramalakshmi et al., 2024 [23]	Gray standard normalization, regional binary patterns	Practical
9	Pereira et al., 2016 [24]	CNN, intensity normalization, bias field correction	Practical
10	Balamurugan & Gnanamanoharan, 2023 [25]	Hybrid FCM-GMM (fuzzy c-means and Gaussian mixture model)	Practical
11	Liu et al., 2021 [26]	U-Net, context-guided attentive CRFs	Practical
12	Lin et al., 2023 [27]	Modality-correlated cross-attention (MCCA), trans and CNN feature calibration (TCFC), transformer	Practical
13	Al-Masni & Kim, 2021 [28]	U-Net, inversion recovery (IR)	Practical
14	Zhang et al., 2021 [29]	GANS for knowledge transfer, cross-modality feature transition (CMFT) process, modalities, and a cross-modality feature fusion (CMFF)	Practical
15	Xu et al., 2023 [30]	Primary feature conservation (PFC) strategy and compact split-attention (CSA)	Practical
16	Peng et al., 2022 [31]	3d U-Net	Practical
17	Ladkat et al., 2022 [32]	3D attention U-Net, a mathematical model for pixel enhancement	Practical
18	Preetha et al., 2021 [33]	CNN	Practical
19	Xu et al., 2022 [34]	CNN, transformer	Practical
20	Telrandhe et al., 2016 [35]	K-means clustering, median filtering, and skull masking	Practical
21	Abd El Kader et al., 2021 [36]	CNN	Practical
22	Zeineldin et al., 2022 [37]	Neuroxai framework	Conceptual
23	Micallef et al., 2021 [38]	U-Net	Practical
24	Athisayamani et al., 2023 [39]	Canny Mayfly algorithm (ACMA), spatial gray level dependence matrix (SGLDM)	Practical
25	Choi et al., 2021 [40]	CNN	Practical
26	Hu et al., 2021 [41]	HPU-Net, fuzzy c-means clustering	Practical
27	Conte et al., 2021 [42]	U-Net	Practical
28	Khan et al., 2022 [43]	CNN	Practical
29	Archana & Komarasamy, 2023 [44]	U-Net	Practical
30	Zhang et al., 2020 [45]	Context residual module, inter-slice context information	Practical
31	Gunasekara et al., 2021 [46]	CNN, Chan–Vese algorithm	Conceptual
32	Anantharajan et al., 2024 [47]	Fuzzy c-means clustering, adaptive contrast enhancement algorithm (ACEA), and a median filter	Practical
33	Chen et al., 2021 [48]	Teacher model with adversarial learning, signed distance maps (SDM)	Practical
34	Gumaei et al., 2019 [49]	The paper does not focus on image segmentation but utilizes a principal component analysis-normalized GIST (PCA-NGIST) feature extraction method without segmentation.	Practical
35	Jia & Chen, 2020 [50]	FAHS-SVM, skull stripping, morphological operations, and wavelet transformation	Practical
36	MS Ullah et al. (2024) [51]	ResNet-50 and Stacked Autoencoders	Practical
37	Tripathy et al., 2023 [52]	Cropping and converting to grayscale, the binary thresholding method	Practical
38	Srinivasan S et al., 2024 [53]	Deep Convolutional Neural Network (CNN)	Practical
39	Tahir et al., 2019 [54]	Otsu method wavelet denoising and histogram equalization	Practical
40	Wadhwa et al., 2019 [55]	The paper reviews multiple approaches, focusing on a hybrid model combining fully convolutional neural networks (FCNN) with conditional random fields (CRF)	Conceptual
41	Farchi, 2023 [56]	The paper focuses on classification, not segmentation, but employs image preprocessing techniques such as resizing, grayscale conversion, image smoothing, and enhancement prior to classification	Practical
42	Chaudhary et al., 2020 [57]	Binary segmentation	Practical
43	Ramamoorthy et al., 2022 [58]	Otsu method	Practical
44	Kiranmayee et al., 2016 [59]	Fuzzy c-means clustering, thresholding method, Watershed segmentation	Practical
45	Lavanya & Nagasundaram, 2023 [60]	Filters	Practical
46	Harish & Ahammed, 2019 [61]	The segmentation aspect is not the focus of the paper. Instead, it discusses image enhancement techniques to improve the quality of brain MRI images	Practical
47	Poornachandra & Naveena, 2017 [62]	N4ITK algorithm	Conceptual
48	Archa & Kumar, 2018 [63]	CNN, wavelet transform, median filtering	Practical
49	Ali et al., 2022 [64]	U-Net, Markov random field (MRF)	Practical
50	Russo et al., 2022 [65]	CNN, transforming Cartesian coordinates into spherical coordinates	Practical
51	Biratu et al., 2021 [66]	CNN	Conceptual
52	Kapoor & Thakur, 2017 [67]	K-means clustering, Fuzzy C-Means clustering, Genetic Algorithms, thresholding method, Watershed segmentation	Conceptual
53	Aggarwal et al., 2023 [68]	Resnet	Practical

, which presents the classification of the 53 articles.

The practical contribution category represents the majority, with 47 articles, while the conceptual contribution category consists of only six studies. This suggests that researchers focus more on developing practical applications than conceptual or theoretical discussions.

3.6.2. Separating Articles That Address Segmentation

After classifying the articles, those specifically focused on brain tumor segmentation were identified, as this was a key aspect of the second research question. The selection process involved examining each article’s title, abstract, and methodology.

As a result, 43 out of 53 articles were selected for further analysis. Figure 12 illustrates the percentage of articles focused on brain tumor segmentation compared to those addressing other research topics.

3.6.3. Studying the Train/Test Mechanism, Evaluation Metrics, and Segmentation Approaches

At this stage, the articles focused on brain tumor segmentation were examined. The primary focus was training and testing techniques, evaluation metrics, and segmentation approaches. These aspects were extracted from the abstracts and methodologies of the selected studies.

The findings reveal various training and testing methods, as well as different evaluation metrics, depending on the role of segmentation. These results are summarized in

Table 4. Summary of training/testing techniques, evaluation metrics, and segmentation approaches in brain tumor segmentation studies.

No.	Authors	TT Mechanism	Evaluation Metrics	Segmentation Approach
1	Díaz-Pernas et al., 2021 [17]	Split 80-20, cross-validation method	DICE score, predicted tumor type accuracy score, and sensitivity	CNN, sliding mechanism
2	Ghafourian et al., 2023 [18]	Split 70-30, cross-validation method	Accuracy, sensitivity, specificity, F1 score	Social spider optimization (SSO) algorithm
3	Menze et al., 2021 [19]	Augmentation	Hausdorff distance, volumetric mismatch	U-Net, intensity normalization, spatial harmonization
4	Kurdi et al., 2023 [20]	MATLAB TOOL,	Pixel accuracy, accuracy, sensitivity, specificity, error rate	Fuzzy c-means clustering
5	Vankdothu & Hameed, 2022 [5]	Split 70-30	DICE score, accuracy, Sensitivity, specificity	K-means clustering, Gray level co-occurrence matrix (GLCM)
6	Daimary et al., 2020 [21]	Split 60-40	DICE score, accuracy	CNN, U-Net
7	Reddy et al., 2018 [22]		DICE score, accuracy, precision, recall, true positive (TP), true negative (TN), false positive (FP), false negative (FN)	K-means clustering
8	Ramalakshmi et al., 2024 [23]		False classification ratio, accuracy, Sensitivity, specificity	Gray standard normalization, regional binary patterns
9	Pereira et al., 2016 [24]	Cross-validation method	DICE score, Sensitivity, Positive Predictive Value (PPV)	CNN, intensity normalization, bias field correction
10	Balamurugan & Gnanamanoharan, 2023 [25]	Split 70-30	Accuracy, sensitivity, specificity, precision, recall, F-score, DICE Similarity Index (DSI)	Hybrid FCM-GMM (fuzzy c-means and Gaussian mixture model)
11	Liu et al., 2021 [26]	Cross-validation method	Sensitivity, specificity, Hausdorff 95 distance	U-Net, context-guided attentive CRFs
12	Lin et al., 2023 [27]	Split 80-10-10	DICE score, sensitivity, Hausdorff 95 distance	Modality-correlated cross-attention (MCCA), Trans and CNN feature calibration (TCFC), transformer
13	Al-Masni & Kim, 2021 [28]	Split 80-20	DICE score, accuracy, sensitivity, specificity, Jaccard index, Matthews correlation coefficient (MCC), and area under the curve (AUC)	U-Net, inversion recovery (IR)
14	Zhang et al., 2021 [29]	Split 80-20	DICE, score, sensitivity, specificity, Hausdorff 95 distance	Gans for knowledge transfer, cross-modality feature Transition (CMFT) process, modalities, and a cross-modality feature fusion (CMFF)
15	Xu et al., 2023 [30]	Split 70-10-20, augmentation, Adam optimization	DICE score, accuracy, precision, recall, mean intersection over union (MIOU)	Primary feature conservation (PFC) strategy and compact split-attention (CSA)
16	Peng et al., 2022 [31]	Split 80-20, augmentation	DICE score, ICC	3D U-Net
17	Ladkat et al., 2022 [32]	Split 77-23	Sensitivity, specificity, Hausdorff95 distance	3D attention U-Net, a mathematical model for pixel enhancement
18	Preetha et al., 2021 [33]	Split 8-0-20, cross-validation method	Dice score, C-index, SSIM	CNN
19	Xu et al., 2022 [34]	Adam optimization	Hausdorff 95 distance, mean Intersection over union (MIOU)	CNN, transformer
20	Telrandhe et al., 2016 [35]	SVM	Accuracy	K-means clustering, median filtering, and skull masking
21	Abd El Kader et al., 2021 [36]	Split 8-0-20, cross-validation method	DICE score, accuracy, sensitivity, specificity, F1 score, precision	CNN
22	Zeineldin et al., 2022 [37]	Augmentation, Z-score	DICE score	Neuroxai framework
23	Micallef et al., 2021 [38]	Split 80-20, cross-validation method, augmentation	DICE score, sensitivity, specificity, Hausdorff95 distance	U-Net
24	Athisayamani et al., 2023 [39]	Split 70-15-15, augmentation	Accuracy, sensitivity, specificity, and recall	Canny Mayfly algorithm (ACMA), spatial gray level dependence matrix (SGLDM)
25	Choi et al., 2021 [40]		Accuracy, precision, and recall	CNN
26	Hu et al., 2021 [41]	Cross-validation method	DICE score, Jaccard coefficient	HPU-Net, fuzzy c-means clustering
27	Conte et al., 2021 [42]	Split 64-20-16, cross-validation method	DICE score	U-Net
28	Khan et al., 2022 [43]	Split 87-13, cross-validation method	Accuracy, error rate	CNN
29	Archana & Komarasamy, 2023 [44]	Split 80-20	DICE score, accuracy, F1 score, precision, recall	U-Net
30	Zhang et al., 2020 [45]	Split 75-0-25, cross-validation, augmentation	Hausdorff95 distance	Context residual module, inter-slice context information
31	Gunasekara et al., 2021 [46]	Split 80-0-20, cross-validation	DICE score, peak signal-to-noise ratio	CNN, Chan–Vese algorithm
32	Anantharajan et al., 2024 [47]	SVM	Accuracy, sensitivity, specificity, peak signal-to-noise ratio, Jaccard coefficient (JC)	Fuzzy c-means clustering, Adaptive Contrast Enhancement Algorithm (ACEA), and a median filter
33	Chen et al., 2021 [48]	Split 80-0-20, cross-validation method	False positive rate (FPR), true positive rate, positive predictive value (PPV)	Teacher model with adversarial learning, signed distance maps (SDM)
34	Jia & Chen, 2020 [50]	SVM	Accuracy, sensitivity, specificity	FAHS-SVM, skull stripping, morphological operations, and wavelet transformation
35	Tripathy et al., 2023 [52]	Split 67-13-20	DICE score, accuracy, sensitivity, specificity, F1 score, precision	Cropping and converting to grayscale, binary thresholding
36	Srinivasan S et al., 2024 [53]	5-fold cross-validation Dataset split into 60 training, 20 validation, 20 testing	Accuracy, Sensitivity, Specificity, Precision, ROC-AUC	classification task only (binary, multi-type, and tumor grade)
37	Tahir et al., 2019 [54]	Split 90-0-10, cross-validation method, SVM	Accuracy, sensitivity, specificity	Wavelet denoising and histogram equalization
38	Ramamoorthy et al., 2022 [58]		Accuracy, specificity, precision	Otsu method
39	Kiranmayee et al., 2016 [59]	SVM	DICE score, accuracy	Fuzzy c-means clustering, thresholding method, Watershed segmentation
40	Archa & Kumar, 2018 [63]	Augmentation	DICE score, accuracy	CNN, wavelet transform, median filtering
41	Ali et al., 2022 [64]		Accuracy, sensitivity, precision	UNET, Markov Random Field (MRF)
42	Russo et al., 2022 [65]	Split 80-20, cross-validation method, augmentation	DICE score, sensitivity, specificity, Hausdorff 95 distance	CNN, transforming Cartesian coordinates to spherical coordinates, Canny
43	Aggarwal et al., 2023 [68]	Split 34-46	Sensitivity, specificity, Jaccard coefficient, peak signal-to-noise ratio, mean square error (MSE)	Resnet

.

4. Discussion

The review reveals that practical contributions dominate the field, with 47 studies focusing on practical applications, compared to only six studies classified as conceptual contributions. This trend highlights a strong inclination among researchers over the past decade toward advancing practical aspects of the field rather than theoretical frameworks.

Furthermore, this review identified two pivotal works [32,44] that introduced groundbreaking methods widely recognized as innovations. These studies have garnered substantial citations, underscoring their significant impact and contributions to the field.

The training and testing mechanisms discussed in this review incorporate various methods and techniques. The most frequently used approach is the one-split method, employed in 26 papers. This method involves dividing the dataset into two parts, typically allocating the majority for training and the remainder for testing (e.g., 80–20% or 70–30% splits). Dataset size and model complexity often determine the split ratio.

The second most common approach is data augmentation, which has been applied in 20 papers to increase dataset size and diversity by introducing transformations to the existing data. The validation technique, used in 14 papers, divides datasets into three parts for training, validation, and testing.

Additional techniques include the following:

• Support vector machines (SVM), noted in five papers;
• MATLAB, Keras, and Z-score normalization.

These methods were employed in 43 papers that utilized segmentation approaches to evaluate model performance. This review highlights the popularity of techniques such as dataset splitting, augmentation, and validation. They are essential steps in deep learning workflows, each with specific advantages and limitations. Splitting the dataset into training, validation, and test sets helps assess model performance objectively, but improper splitting may lead to biased results. Data augmentation enhances model generalization by creating diverse samples from limited data; however, excessive or unrealistic transformations can introduce noise. Validation techniques, such as cross-validation, improve reliability but increase computational cost. A balanced approach is required to ensure robust and meaningful model evaluation, which is widely adopted for its effectiveness in improving results.

The evaluation metrics summarized in

Table 5. Distribution of metrics in the articles.

No.	Authors	Accuracy	Sensitivity	Specifity	Recall OR (TPR)	F1 Score	Precision OR (ppv)	Dice Score	Hausdorff 95 Distance	Other Classification Metrics	Other Segmentation Metrics
1	Díaz-Pernas et al., 2021 [17]	X	X					X
2	Ghafourian et al., 2023 [18]	X	X	X		X
3	Menze et al., 2021 [19]										X
4	Kurdi et al., 2023 [20]	X	X	X						X	X
5	Vankdothu & Hameed, 2022 [5]	X	X	X				X
6	Daimary et al., 2020 [21]	X						X
7	Reddy et al., 2018 [22]	X			X		X	X		X
8	Ramalakshmi et al., 2024 [23]	X	X	X						X
9	Pereira et al., 2016 [24]		X				X	X
10	Balamurugan & Gnanamanoharan, 2023 [25]	X	X	X	X	X	X	X
11	Liu et al., 2021 [26]		X	X					X
12	Lin et al., 2023 [27]		X	X					X
13	Al-Masni & Kim, 2021 [28]	X	X	X				X		X
14	Zhang et al., 2021 [29]		X	X				X	X
15	Xu et al., 2023 [30]	X			X		X	X			X
16	Peng et al., 2022 [31]							X			X
17	Ladkat et al., 2022 [32]		X	X					X
18	Preetha et al., 2021 [33]							X		X	X
19	Xu et al., 2022 [34]								X		X
20	Telrandhe et al., 2016 [35]	X
21	Abd El Kader et al., 2021 [36]	X	X	X		X	X
22	Zeineldin et al., 2022 [37]							X
23	Micallef et al., 2021 [38]		X	X				X	X
24	Athisayamani et al., 2023 [39]	X	X	X	X
25	Choi et al., 2021 [40]	X			X		X
26	Hu et al., 2021 [41]							X			X
27	Conte et al., 2021 [42]							X
28	Khan et al., 2022 [43]	X								X
29	Archana & Komarasamy, 2023 [44]	X			X	X	X	X
30	Zhang et al., 2020 [45]								X
31	Gunasekara et al., 2021 [46]							X			X
32	Anantharajan et al., 2024 [47]	X	X	X							X
33	Chen et al., 2021 [48]				X		X			X
34	Jia & Chen, 2020 [50]	X	X	X
35	Tripathy et al., 2023 [52]	X	X	X		X	X	X
36	Srinivasan S et al., 2024 [53]	X	X	X			X			X
37	Tahir et al., 2019 [54]	X	X	X
38	Ramamoorthy et al., 2022 [58]	X		X			X
39	Kiranmayee et al., 2016 [59]	X						X
40	Archa & Kumar, 2018 [63]	X						X
41	Ali et al., 2022 [64]	X	X				X
42	Russo et al., 2022 [65]		X	X				X	X
43	Aggarwal et al., 2023 [68]		X	X						X	X

demonstrate the diversity of approaches used to assess model performance. The DICE score and accuracy are indeed popular metrics because of their straightforward application and ability to provide immediate insights into model performance. Among these, the accuracy score, cited in 25 papers, measures the ratio of correctly identified instances (true positives and true negatives) relative to the total number of instances. The DICE score, used in 23 papers, evaluates the overlap between predicted and actual sets, making it a crucial metric for image segmentation and binary classification tasks. Sensitivity, sometimes referred to as recall, measures the proportion of actual positives that are correctly identified by the model. Specificity measures the proportion of actual negatives that are accurately identified as such, thus avoiding false positives. F1-score is the harmonic mean of precision (the number of true positive results divided by the number of all positive results, including those not relevant) and recall.

Additional metrics such as true positives (TP), true negatives (TN), false positives (FP), false negatives (FN), false positive rate (FPR), negative predictive value (NPV), boundary displacement error (BDE), mean intersection over union (mIoU), false classification rate, pixel accuracy, mismatch, volumetric overlap, Matthews correlation coefficient (MCC), and area under the curve (AUC) further illustrate the wide range of evaluation methods utilized in the literature.

The variety of evaluation metrics reflects the demands of different tasks and datasets. While accuracy remains a widely used metric, its limitations in addressing imbalanced datasets necessitate supplementary measures such as sensitivity, specificity, and F1 score for a more comprehensive analysis. In image segmentation tasks, metrics like the DICE score, Jaccard coefficient, and Hausdorff95 distance provide deeper insights into model performance. Domain-specific metrics such as PSNR and MSE are also critical in evaluating image quality in relevant contexts.

The selection of appropriate metrics should align with the research objectives, dataset characteristics, and application requirements to ensure thorough and meaningful assessments of model efficacy.

GANs, while powerful generative models, have faced several criticisms that can limit their practical application, especially when compared to more established architectures like CNNs and U-Nets for specific tasks like image segmentation.

Emerging techniques like federated learning and self-supervised learning hold significant promise in the realm of tumor classification, yet they remain relatively underexplored. Federated learning offers a solution to privacy concerns by allowing multiple institutions to collaboratively train models without the need to share sensitive patient data.

On the other hand, self-supervised learning is valuable in situations where annotated data are scarce. It allows models to extract meaningful patterns and representations from large volumes of unlabeled data.

5. Conclusions

This systematic review analyzed literature on image processing, particularly medical images, focusing on brain tumor detection using magnetic resonance imaging (MRI). Fifty-three articles from the past ten years were reviewed and selected for their relevance to brain tumor detection. The analysis showed that 43 of the 53 studies employed brain tumor segmentation techniques, and 47 included practical applications of their proposed methods. Most studies relied on convolutional neural networks (CNNs) and machine learning (ML) models. Neural networks are recognized for their strong learning ability and accuracy, but require large datasets. Fuzzy logic offers interpretability and handles uncertainty well, but may lack precision in complex patterns. Genetic algorithms are effective in feature selection and optimization, but are computationally intensive. By comparing these techniques, the review provides insights into their suitability for different tumor classification tasks.

However, several limitations were identified, including long training times for datasets, particularly in cases involving hybrid models or complex ML algorithms. Additionally, some studies had insufficient dataset sizes, limiting their ability to achieve optimal accuracy. These findings emphasize the need for further in-depth research to enhance accuracy, as the results directly impact patient outcomes. Therefore, several approaches can be adopted to enhance the model’s accuracy when dealing with limited medical image data, like (data augmentation, transfer learning, synthetic data generation, and others).

Prioritizing tumor segmentation, localization, and classification remains critical to improving the precision and reliability of brain tumor detection methods. Future studies should explore larger, more diverse datasets and develop optimized computational strategies to mitigate processing inefficiencies. Additionally, integrating multi-modal imaging techniques and leveraging advanced AI models could drive significant breakthroughs in early tumor detection, ultimately leading to improved clinical decision-making and patient survival rates.

Ethical Considerations

While this study does not directly utilize patient data, it is essential to emphasize the ethical standards pertinent to research of this nature. Upholding patient privacy and adhering to data protection regulations like GDPR or HIPAA is a crucial obligation in clinical and health data research. The articles utilized for this study had obtained necessary ethical approval, ensuring compliance with institutional and regulatory guidelines. Any future work involving direct data collection will explicitly address obtaining ethical approvals as well as establishing protocols for handling such information.