Optimizing Glioblastoma Multiforme Diagnosis: Semantic Segmentation and Survival Modeling Using MRI and Genotypic Data

Abstract

Glioblastoma multiforme (GBM) is the most aggressive and common primary brain tumor. Magnetic resonance imaging (MRI) provides detailed visualization of tumor morphology, edema, and necrosis. However, manually segmenting GBM from MRI scans is time-consuming, subjective, and prone to inter-observer variability. Therefore, automated and reliable segmentation methods are crucial for improving diagnostic accuracy. This study employs an image semantic segmentation model to segment brain tumors in MRI scans of GBM patients. The MRI recall images include T1-weighted imaging (T1WI) and fluid-attenuated inversion recovery (FLAIR) sequences. To enhance the performance of the semantic segmentation model, image preprocessing techniques were applied before analyzing and comparing commonly used segmentation models. Additionally, a survival model was constructed using discrete genotype attributes of GBM patients. The results indicate that the DeepLabV3+ model achieved the highest accuracy for semantic segmentation, with an accuracy of 77.9% on T1WI image sequences, while the U-Net model achieved 80.1% accuracy on FLAIR image sequences. Furthermore, in constructing the survival model using a discrete attribute dataset, the dataset was divided into three subsets based on different missing value handling strategies. This study found that replacing missing values with 1 resulted in the highest accuracy, with the Bernoulli Bayesian model and the multinomial Bayesian model achieving an accuracy of 94.74%. This study integrates image preprocessing techniques and semantic segmentation models to improve the accuracy and efficiency of brain tumor segmentation while also developing a highly accurate survival model. The findings aim to assist physicians in saving time and facilitating preliminary diagnosis and analysis.

1. Introduction

Glioblastoma multiforme (GBM), also known simply as glioblastoma, is the most aggressive and most common primary brain tumor in adults. GBM originates from astrocytes, the star-shaped glial cells in the brain and spinal cord, and is classified by the World Health Organization as a Grade IV glioma—a designation reserved for the most malignant tumors [1]. GBM accounts for 48.6% of all malignant tumors in the central nervous system. This high percentage highlights GBM as the most prevalent central nervous system cancer [2]. The median overall survival for GBM patients is low at just 15 months [3]. Malignant gliomas (with GBM being a major subset) are responsible for 2.5% of all cancer-related deaths, indicating GBM’s broader impact on public health [4].

Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that uses a strong magnetic field and radio waves to generate signals from hydrogen atoms within the body. These signals are then processed by a computer to produce high-resolution images of organs and tissues. In the diagnostic process, MRI plays a crucial role in establishing a differential diagnosis during a patient’s initial consultation. It can be used to accurately identify the location and size of brain tumors. MRI scans can reveal characteristics of GBM, and commonly used MRI sequences include T1-weighted imaging (T1WI), T2-weighted imaging, fluid-attenuated inversion recovery (FLAIR), and T1-weighted contrast-enhanced [5,6]. The T1WI is a sequence that primarily reflects the T1 relaxation times of tissues. This sequence provides basic anatomical information and helps distinguish between different brain tissues. The T2-weighted imaging, on the other hand, is based on T2 relaxation times. It is sensitive to water content, making edema and certain parts of the tumor appear bright. FLAIR is a special type of T2-weighted imaging with an added inversion recovery pulse that suppresses free fluid signals, especially from cerebrospinal fluid. FLAIR is particularly valuable for segmenting edematous regions, which often contain infiltrative tumor cells beyond the enhancing core. T1-weighted contrast-enhanced adds a gadolinium-based contrast agent that shortens the T1 relaxation time of tissues where it accumulates. It shows strong enhancement in the tumor core region.

Image segmentation, the process of dividing an image into meaningful regions, plays a crucial role in the quantitative analysis of brain tumors. Specifically, segmentation of brain tumor MRI involves identifying and delineating tumor boundaries, which is essential for measuring tumor volume, monitoring growth over time, and planning surgical interventions. However, manual segmentation by radiologists is time-consuming, subjective, and prone to inter-observer variability, especially given the complexity and heterogeneity of brain tumors [7]. These limitations highlight the need for automated, robust, and accurate segmentation methods to assist clinicians in making informed decisions.

Early efforts in brain tumor segmentation relied heavily on traditional image processing techniques, such as thresholding, region growing, and clustering [8]. While these methods offer simplicity and interpretability, they are often sensitive to noise, artifacts, and heterogeneous appearances in MRI scans, resulting in inconsistent performance across datasets. To overcome the limitations of traditional methods, researchers turned to machine learning approaches, which offer more adaptive and data-driven solutions. These methods typically involve two stages: feature extraction and classification. Features such as texture, intensity, and spatial information are extracted from MRI images and fed into classifiers like support vector machine (SVM) or random forest (RF). For instance, Bahadure et al. [9] proposed an SVM-based method using texture and intensity features for tumor segmentation, achieving significant improvements over traditional methods. However, these techniques require elaborate feature engineering, which is both time-consuming and dependent on expert knowledge. Moreover, handcrafted features may fail to capture the complex patterns present in brain tumor images, limiting their generalization capability.

With the rise of deep learning, particularly following the success of AlexNet [10] in 2012 in image classification tasks, convolutional neural networks (CNNs) began to be applied in medical image analysis. Unlike earlier approaches, CNNs can automatically learn hierarchical features directly from raw image data, eliminating the need for manual feature design. In 2015, the introduction of U-Net [11] marked a milestone in brain tumor segmentation. U-Net, specifically designed for medical image segmentation, uses an encoder–decoder architecture with skip connections, enabling it to effectively capture fine-grained features. U-Net has become the cornerstone of brain tumor segmentation and has inspired numerous adaptations.

The evolution of medical image segmentation has progressed from manual feature design to traditional machine learning-based automated classification and, ultimately, to deep learning methods featuring end-to-end automatic feature learning and segmentation. Although the emergence of deep learning has significantly improved the accuracy and automation of segmentation tasks, its performance still varies across different medical imaging datasets. Traditional methods continue to hold value in specific applications or when integrated as part of hybrid approaches. The application of image processing techniques can enhance the performance of deep learning models.

The application of image processing techniques can enhance the performance of deep learning models. Ullah et al. [12] evaluated various preprocessing methods, including bias field correction (BFC), intensity normalization, histogram equalization, and Gibbs ringing artifact removal (GRAR), to assess their impact on tumor segmentation in brain MRI images. The results showed that these preprocessing techniques significantly improved the segmentation accuracy of the 3D U-Net model. Hussein et al. [13] proposed using contrast limited adaptive histogram equalization (CLAHE) to enhance the contrast of X-ray images, followed by classification using a CNN. Their approach effectively increased the accuracy of lung disease classification. Similarly, Marciniak et al. [14] assessed the impact of CLAHE on the classification of retinal abnormalities in optical coherence tomography images. Their findings indicated that CLAHE preprocessing enhanced deep learning classification accuracy, with improvements of up to 4.75%.

Therefore, the objective of this study is to enhance the segmentation performance of deep learning models on GBM MRI scans by integrating image processing techniques with deep learning architectures. The methodology involves using MRI sequences as the dataset. Preprocessing is performed using several image enhancement techniques, including adaptive histogram equalization (AHE), intensity normalization, BFC, and GRAR. Subsequently, deep learning segmentation models (U-Net, U-Net++, DeepLabV3+, and LinkNet) are applied to segment brain tumors. The accuracy of these models is evaluated to assist clinicians in making more effective preliminary diagnoses.

Furthermore, the study incorporates discrete patient attribute data and applies machine learning techniques, including decision tree (DT), RF, extreme gradient boosting (XGBoost), SVM, multi-layer perceptron (MLP), K-nearest neighbors (KNN), and naive Bayes (NB), to analyze gene expression and anti-cancer drug-related attributes, aiming to predict patient survival outcomes and support clinical decision-making.

The primary contributions of this study are delineated as follows:

• This study explores the combination of image processing techniques and deep learning models for segmenting GBM MRI images and finds that such integration can significantly improve segmentation performance.
• For T1WI, the highest Dice coefficient of 0.779 is achieved when applying Intensity Normalization as preprocessing followed by the DeepLabV3+ model. For FLAIR imaging, the highest Dice coefficient of 0.801 is obtained when using GRAR preprocessing combined with the U-Net model.
• The DeepLabV3+ model shows Dice coefficient improvements of 0.156 and 0.149 on T1WI and FLAIR images, respectively, when using intensity normalization as preprocessing.
• Using gene data from GBM patients, this study predicts survival outcomes. When missing values are replaced with 1, the NB model achieves an accuracy of 0.9474.

2. Related Works

2.1. Medical Image Features

Medical image features refer to quantifiable attributes extracted from imaging modalities such as MRI, computed tomography, X-ray, or ultrasound. These features (such as color, texture, shape, and grayscale intensity) are essential for assisting clinicians in tasks like tumor identification and tissue structure evaluation. Prior studies have demonstrated that such features play a crucial role in medical image analysis, typically categorized into intensity, shape, and texture features [15]. Intensity features describe the distribution of pixel values in the image and are often used to differentiate tissues or detect lesions. Metrics like mean, maximum, minimum, standard deviation, skewness, and kurtosis are frequently used to summarize pixel intensity [16]. Shape features focus on the geometric properties and boundary characteristics of lesions or organs, providing information such as volume, surface area, aspect ratio, and boundary irregularity [17]. Texture features capture the spatial distribution and regularity of grayscale values, aiding in distinguishing between different tissue types. Common methods for extracting texture features include the gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, and wavelet transform analysis [17,18].

These features can be extracted manually (referred to as hand-crafted features) or learned automatically using deep learning models. With the advancement of deep learning, especially CNNs, it has become possible to automatically learn more abstract and discriminative features from large-scale datasets. These deep learning-based features, which emerge through layers of convolution and pooling, may be less interpretable but tend to outperform traditional features in complex segmentation tasks.

2.2. Medical Image Segmentation

Historically, due to computational limitations, early medical image segmentation methods relied on basic techniques such as thresholding, region growing, edge detection, and clustering [8]. Thresholding segments images based on pixel intensity, while region growing starts from a seed point and groups neighboring pixels with similar properties. Edge detection methods like the Sobel and Canny algorithms identify sharp intensity changes that indicate boundaries, and clustering approaches such as K-means group pixels based on similarity in intensity or position.

With the emergence of machine learning, segmentation shifted toward model-based classification using features extracted from image regions or pixels. Classifiers such as SVM and RF were commonly applied to predict pixel labels. These traditional machine learning approaches had advantages in interpretability and required less training data, but they were limited by their reliance on manually designed features and their difficulty in handling complex or highly variable image content.

Since 2012, deep learning has revolutionized medical image segmentation, largely replacing traditional methods due to its superior performance and automation capabilities. Deep learning models can learn hierarchical, high-level features directly from raw image data without the need for manual design. For instance, in the application of deep learning methods to retinal vessel segmentation, the presence of numerous capillaries and complex tree-like structures in retinal vessels makes it difficult for traditional segmentation techniques to perform effectively [19]. Therefore, deep learning approaches are necessary to improve segmentation accuracy in such cases. Luu et al. [20] developed a deep learning model based on an improved U-Net architecture for brain tumor segmentation in MRI images, achieving top performance in the Brain Tumor Segmentation 2021 challenge. This demonstrates how modern deep learning approaches have become the dominant technique in medical image segmentation, delivering substantial improvements over classical methods. Ali et al. [21] investigated the application of deep learning in diagnosing neurological disorders using brain MRI images. The results demonstrated that deep learning models outperformed traditional machine learning methods in terms of classification accuracy, especially when dealing with high-dimensional and complex feature data.

2.3. Deep Learning for Brain Tumor Image Segmentation

The application of deep learning in medical image segmentation has become a research hotspot in recent years, especially for processing complex medical images such as MRI, computed tomography, and X-rays. Fully convolutional networks (FCNs) [22], introduced in 2015, replaced the fully connected layers in traditional CNNs with convolutional layers, enabling end-to-end pixel-level prediction. This means that the output of an FCN is a segmentation map of the same size as the input image, where each pixel is assigned a class label. Later, U-Net was developed as an improvement upon the FCN and became particularly suitable for high-precision segmentation in medical imaging.

Chen et al. [23] and Noori et al. [24] used U-Net for brain tumor image segmentation. In addition, Islam and colleagues [25] not only used U-Net for segmenting brain tumor images but also employed it to predict patient survival. These studies demonstrate that U-Net is highly suitable for brain tumor image segmentation. The name “U-Net” originates from the U-shaped architecture formed by its encoder and decoder paths. This design has proven particularly effective in medical image segmentation and other applications requiring precise pixel-level classification.

In the related literature, Cui et al. [26] applied U-Net++ [27] for brain tumor segmentation in CT images. Wu et al. [28] used U-Net++ for the automatic segmentation of pelvic organs after hysterectomy, while Micallef, Seychell, and others [29] utilized U-Net++ for automated brain tumor image segmentation. These studies show that U-Net++ is an effective model for segmenting brain tumor MRI images.

Roy Choudhury et al. [30] employed DeepLabV3+ [31] for brain tumor image segmentation. Wang [32] used both DeepLabV3+ and U-Net for leaf image segmentation and classification. Meanwhile, Liu, Zhang, and colleagues [33] used DeepLabV3+ in a study on bladder cancer image segmentation. These examples suggest that DeepLabV3+ is a viable model for brain tumor image segmentation.

Regarding research on LinkNet [34], Sobhaninia et al. [35] and Naz et al. [36] adopted LinkNet as the model architecture for analyzing brain tumor images. Hemanth and colleagues [37] used LinkNet for brain tumor object detection. These findings indicate that the LinkNet model is also applicable for brain tumor image segmentation.

2.4. Digital Image Processing

One of the primary tasks in digital image processing is image enhancement and restoration, as digital images are often affected by noise, blurring, and other distortions. Image processing techniques are therefore particularly crucial. Given the widespread application of digital imaging in medical contexts, image processing plays an essential and irreplaceable role in improving image quality [38].

AHE [39] is a computer image processing technique used to enhance local contrast in images. It works by computing histograms of specified regions and redistributing the image brightness values based on those histograms, thereby modifying the image’s contrast. This method is particularly suited for improving local contrast and enhancing image edges to reveal more details. While AHE can also reduce noise, it may degrade the visual quality of MRI images.

Intensity normalization [40] is a method that maps image intensity values into a standardized, predefined range. During the MRI acquisition process, different scanners or scanning parameters may be used at different times to scan different subjects or even the same subject. This can lead to significant variations in intensity, which severely affect MRI image consistency. To mitigate such variability, intensity normalization is applied. This method leverages geometric moments of images to find a set of parameters that eliminate the effects of various transformations, converting the original raw image into a unique standard form. This often involves affine transformations such as translation, rotation, and scaling, to normalize images to the same dimensional scale or equivalent intensity range.

BFC [41] is an image processing technique primarily used to correct intensity inhomogeneity in images. In medical imaging (especially MRI) non-uniform brightness may arise due to hardware limitations of the scanner, physiological factors of the patient, or other unknown causes. Such inconsistencies result in the same tissue appearing with different grayscale values across the image. Bias field correction serves as a preprocessing step in MRI to correct these distortions. In this study, the method used is N4ITK, which was proposed by Tustison, to perform bias field correction.

GRAR [42] is a method for eliminating artifacts in MRI images. Gibbs artifacts are typically the result of using the Fourier transform to reconstruct MRI signals into images. These artifacts appear as false oscillations or ripples near regions with abrupt intensity changes, such as edges or interfaces. To eliminate Gibbs ringing from MRI images, the local sub-voxel shift technique is employed.

3. Materials and Methods

The hospital provides data of GBM patients, including MRI and genetic data. The MRI data is filtered and annotated to form the raw dataset, which then undergoes four image processing techniques: AHE, intensity normalization, BFC, and GRAR. Each technique generates a corresponding dataset. The datasets are then subjected to image segmentation using four semantic segmentation models: U-Net, U-Net++, DeepLabV3+, and LinkNet.

Additionally, the genetic data is filtered to create a usable dataset, with missing values replaced by 0, 1, and 2, creating three datasets. These three datasets are then used with various machine learning classification models to predict patient survival. The machine learning models include DT [43], RF [44], XGBoost [45], SVM [46], MLP [47], KNN [48], and NB [49]. NB includes Gaussian naive Bayes (GNB), multinomial naive Bayes (MNB) and Bernoulli naive Bayes (BNB). The research workflow is shown in Figure 1.

3.1. Datasets

The dataset used in this study is sourced from the Taichung Veterans General Hospital in Taiwan (IRB No: CF17263B-3), which provided MRI scans of patients with glioblastoma. The dataset includes MRI scans from 69 patients, with each MRI consisting of T1WI and FLAIR image sequences. Each sequence contains 1695 images, totaling 3390 images.

The inclusion criteria for patients were being at least 20 years of age and having a diagnosis of GBM, regardless of gender. The exclusion criteria included patients under the age of 20, those with concurrent tumors or other specific medical conditions, and members of vulnerable populations (including children, minors, pregnant women, individuals with mental disorders, critically ill patients, and those with behavioral disorders).

The hospital also provided genetic data, including discrete attribute data on the patients’ gender, survival status, use of the anti-cancer drug Avastin, and the presence of genes such as VEGF, IDH, hTERT, MGMT, p53, and p21. A total of 222 patients are included in this dataset.

The following are some genes or factors associated with GBM.

• VEGF (vascular endothelial growth factor) [50] primarily binds to receptors on vascular endothelial cells (VEGF receptors, VEGFR), thereby promoting tumor growth.
• IDH (isocitrate dehydrogenase) [51] refers to a group of enzymes divided into IDH1 and IDH2. Mutations in IDH are associated with the development of glioblastoma.
• hTERT (human telomerase reverse transcriptase) [52] is commonly considered an oncogene when mutated.
• MGMT (O6-methylguanine-DNA methyltransferase) [53] is a DNA repair gene. MGMT is often involved in promoter methylation, which can interfere with the DNA repair process.
• p53 [54] refers to a family of homologous proteins known as tumor suppressors.
• p21 [55], also known as CDKN1A (cyclin-dependent kinase inhibitor 1A), is a tumor suppressor protein involved in regulating the cell cycle.

The presence and status of these GBM-related genes or factors are typically revealed in a patient’s genetic test results.

3.2. Screening and Labeling of MRI Images

The MRI data underwent a preprocessing pipeline, including image selection and tumor annotation, to prepare for model training. MRI scans consist of a series of consecutive slices captured from the top of the head downward. However, not all slices clearly depict brain tumors or associated edema, which are critical for effective model training. To streamline subsequent research, this study manually selected slices that clearly show brain tumors or edema to construct a curated dataset.

The selected MRI slices were annotated using LabelMe, an open-source tool that supports drawing polygons, rectangles, circles, lines, and points. LabelMe offers versatile functionality and can be used for data annotation tasks such as classification, object detection, semantic segmentation, and instance segmentation. In this study, LabelMe was used to manually delineate the contours of brain tumors and edema, producing precise annotations to support downstream research and model development.

3.3. Screening and Handling Missing Values of Genetic Data

In this study, the patients’ tabular data was also filtered prior to building the survival prediction models to ensure data usability. The dataset obtained from the hospital contains information on 222 patients. However, many fields contained missing values. After filtering, the missing values were replaced with 0, 1, and 2, respectively, in order to construct usable datasets for model training.

3.4. Image Processing

Image processing techniques play a crucial and indispensable role in enhancing the quality of medical imaging. In this study, MRI is used as the dataset, including T1WI and FLAIR image sequences. Four image processing techniques (AHE, intensity normalization, BFC, and GRAR) are applied to both T1WI and FLAIR sequences. These preprocessing steps are essential for the subsequent development of image segmentation models.

3.4.1. AHE

AHE divides the image into multiple regions and applies histogram equalization independently to each tile. Afterward, bilinear interpolation is performed to smooth the boundaries between adjacent regions.

The new intensity value $I$ is presented in Formula (1):

$$I\left(i\right)={\displaystyle \frac{CDF\left(i\right)-{CDF}_{min}}{{CDF}_{max}-{CDF}_{min}}}\times \left(L-1\right),$$

(1)

where $i$ is the pixel value of the image, $CDF\left(i\right)$ is the cumulative distribution function (CDF) of pixel value $i$, ${CDF}_{min}$ is the CDF of the minimum pixel value, ${CDF}_{max}$ is the CDF of the maximum pixel value, and $L$ is the total number of pixels.

3.4.2. Intensity Normalization

The calculation method of intensity normalization is as follows (2).

$$I={\displaystyle \frac{X-I_{min}}{I_{max}-I_{min}}},$$

(2)

where $I$ is the new intensity value, $X$ is the original intensity value, $I_{max}$ is the maximum intensity value, and $I_{min}$ is the minimum intensity value.

3.4.3. BFC

The N4ITK algorithm operates on the assumption that the observed MRI image $v$ is modeled as shown in Formula (3):

$$v=uf+n,$$

(3)

where $u$ is the true underlying image without bias, $f$ is the multiplicative bias field, $n$ is additive noise.

Assuming the noise is negligible after log transformation, the model simplifies to shown in Formula (4):

$$\widehat{v}=\widehat{u}+\widehat{f},$$

(4)

where $\widehat{v}$ = $\log v$, $\widehat{u}$ = $\log u$, and $\widehat{f}$ = $\log f.$

The corrected image is then obtained by subtracting the estimated bias field from the observed image, as shown in Formula (5):

$${\widehat{u}}^n={\widehat{v}}^n-{\widehat{f}}^n,$$

(5)

where $n$ denotes the $n$-th iteration.

This process continues until convergence criteria are met, typically when changes between iterations fall below a predefined threshold.

3.4.4. GRAR

The core idea of GRAR is to reduce the Gibbs effect by averaging the image after subtle spatial shifts. This method of local sub-pixel shifting and averaging can effectively eliminate Gibbs artifacts while preserving edge information.

The new intensity value $I$ is presented in Formula (6):

$$I={FT}^{-1}\left\{FT\left\{I_x\right\}·G_x+FT\left\{I_y\right\}·G_y\right\},$$

(6)

where $FT\left\{·\right\}$ denotes the Fourier transform, $I_x$ is the corrected image in the x dimension, $I_y$ is the corrected image in the y dimension, $G_x$ is the weighting function in the x dimension, and$G_y$ is the weighting function in the x dimension.

$G_x$ and $G_y$ are as shown in Formula (7):

$$G_x={\displaystyle \frac{1+\cos k_y}{\left(1+\cos k_y\right)+\left(1+\cos k_x\right)}}, G_y={\displaystyle \frac{1+\cos k_x}{\left(1+\cos k_x\right)+\left(1+\cos k_y\right)}},$$

(7)

where $k_y$ is the $k$ radian of the $y$-axis, and $k_x$ is the $k$ radian of the $x$-axis. When $G_x+G_y=1$, the artifacts can be eliminated.

3.5. Semantic Segmentation

This study employs four semantic segmentation models, including U-Net, U-Net++, DeepLabV3+, and LinkNet.

3.5.1. U-Net

The architecture of U-Net (Figure 2) is structured as a symmetric, it comprises two main components: a contracting path and an expansive path. The contracting path functions similarly to a traditional convolutional network. It involves repeated application of two 3 × 3 convolutions, each followed by a rectified linear unit (ReLU) activation, and then a 2 × 2 max pooling operation with stride 2 for downsampling. As the network goes deeper, the number of feature channels doubles, enabling the model to capture increasingly abstract and complex features. On the other hand, the expansive path mirrors the structure of the contracting path but in reverse. It begins with a 2 × 2 up-convolution (also known as a transposed convolution) that reduces the number of feature channels by half. Each upsampling step is followed by a concatenation with the corresponding feature map from the contracting path, facilitated by skip connections. These skip connections are critical, as they reintroduce spatial information lost during downsampling. After concatenation, the model applies two more 3 × 3 convolutions followed by ReLU activations, continuing the pattern from the encoder. The final layer is a 1 × 1 convolution, which maps the feature representation to the desired number of classes, typically resulting in a pixel-wise segmentation output.

3.5.2. U-Net++

The fundamental idea in U-Net++ (Figure 3) is to gradually evolve feature maps before they are passed to the decoder. Instead of simply copying encoder features to the decoder through a single skip connection, U-Net++ uses intermediate convolution layers that sit between the encoder and decoder blocks. These layers progressively refine the encoder’s output at each spatial resolution, resulting in a series of convolution blocks forming short paths and deeper convolution chains that improve feature alignment. Each of these intermediate nodes receives features not only from its immediate predecessor but also from earlier layers, creating a dense network of connections across the architecture.

3.5.3. DeepLabV3+

The encoder in DeepLabV3+ employs Xception, where atrous convolutions are applied to extract dense features from the input image. To further aggregate multi-scale context, it incorporates the atrous spatial pyramid pooling (ASPP) module. ASPP consists of parallel atrous convolutions with different dilation rates, enabling the model to capture features at various scales simultaneously. This spatial pyramid of filters helps the network become more sensitive to objects of different sizes in a single forward pass, as shown in Figure 4.

3.5.4. LinkNet

What sets LinkNet apart is how it handles the decoding phase. Instead of relying on complex mechanisms to recover spatial resolution, LinkNet uses simple yet effective decoder blocks that are directly linked to their corresponding encoder blocks. This one-to-one linkage is where the model gets its name. The decoder receives two sources of information: the upsampled output from the previous decoder layer and a bypassed, compressed version of the encoder’s output from the same resolution level. These bypass connections are not raw feature maps but are first passed through 1 × 1 convolutions, reducing their channel dimensionality and helping preserve essential spatial features without overwhelming the decoder. Each decoder block then combines the upsampled features with these lightweight encoder features, followed by a few convolution operations to refine the result. This fusion process is repeated at each scale until the spatial resolution is restored to match the original input size. As shown in Figure 5.

4. Results and Discussion

4.1. The Results of Screening and Labeling of MRI Images

The dataset obtained from the hospital included 1695 T1WI sequence images and 1695 FLAIR sequence images from 69 patients. Through manual curation and selection, we compiled a usable dataset by retaining only the images that clearly captured brain tumors or edema. As a result, the final dataset consists of 433 T1WI sequence images and 433 FLAIR sequence images from the same 69 patients.

4.2. The Results of Screening and Handling Missing Values of Genetic Data

The genetic data obtained from the hospital originally contained data from 222 patients. If a record had missing values in more than two fields, it was excluded from the dataset. As a result, data from only 61 patients were retained after filtering. The missing values were replaced with 0, 1, and 2, where 0 indicates absence, 1 indicates presence, and 2 indicates unknown. As a result, three gene datasets were created: Gene_0, Gene_1, and Gene_2.

4.3. Image Processing Results

This study utilized MRI image sequences of brain tumor patients, including T1WI and FLAIR sequences. To enhance the performance of subsequent model training, in addition to using the original images, four types of image preprocessing were applied to the MRI sequences: AHE, intensity normalization, BFC, and GRAR.

For the T1WI sequence, the original image dataset is referred to as the T1_OG dataset. After applying the four preprocessing methods to the original dataset, the following datasets were generated: T1_AHE, T1_Norm, T1_BFC, and T1_GRAR. For the FLAIR sequence, the original image dataset is referred to as the FLAIR_OG dataset. Similarly, after applying the four preprocessing methods, the following datasets were generated: FLAIR_AHE, FLAIR_Norm, FLAIR_BFC, and FLAIR_GRAR.

4.4. Semantic Segmentation Results

In this study, four semantic segmentation models (U-Net, U-Net++, DeepLabV3+, and LinkNet) were used. The hyperparameters were set as follows: batch size = 8, epochs = 50, optimizer = “Adam”, learning rate = 1 × 10⁻⁴, and the loss function was defined as a combination of 50% binary cross-entropy loss and 50% Dice loss. The dataset was split into training, validation, and testing sets in a 5:2:3 ratio. Evaluation metrics included Dice coefficient, intersection over union (IoU), recall, and precision. Each image in the testing set was evaluated, and the results were averaged to obtain the final performance scores.

4.4.1. Comparison Results of Model Performance for T1WI Sequence

The four semantic segmentation models were tested on five different T1WI datasets. The results are presented in

Table 1. The test results on T1WI images.

Dataset	Model	Dice	IoU	Recall	Precision
T1_OG	U-Net	0.609	0.438	0.593	0.625
	U-Net++	0.506	0.339	0.532	0.483
	DeepLabV3+	0.623	0.452	0.599	0.648
	LinkNet	0.427	0.271	0.407	0.449
T1_AHE	U-Net	0.631	0.461	0.618	0.644
	U-Net++	0.525	0.356	0.501	0.552
	DeepLabV3+	0.604	0.433	0.622	0.587
	LinkNet	0.424	0.269	0.402	0.448
T1_Norm	U-Net	0.709	0.549	0.701	0.717
	U-Net++	0.583	0.411	0.591	0.575
	DeepLabV3+	0.779	0.638	0.749	0.812
	LinkNet	0.427	0.271	0.410	0.445
T1_BFC	U-Net	0.655	0.487	0.638	0.673
	U-Net++	0.519	0.350	0.493	0.548
	DeepLabV3+	0.684	0.52	0.693	0.675
	LinkNet	0.426	0.271	0.405	0.449
T1_GRAR	U-Net	0.681	0.516	0.702	0.661
	U-Net++	0.538	0.368	0.542	0.534
	DeepLabV3+	0.711	0.552	0.748	0.678
	LinkNet	0.428	0.272	0.443	0.414

.

As shown in Table 1, the test results on the T1_OG dataset indicate that the DeepLabV3+ model achieved the highest Dice coefficient, reaching 0.623. On the T1_AHE dataset, the U-Net model performed best, with a Dice coefficient of 0.631. For the T1_Norm dataset, the DeepLabV3+ model again achieved the highest Dice coefficient, reaching 0.779. On the T1_BFC dataset, the DeepLabV3+ model outperformed the others, with a Dice coefficient of 0.684. Among all the T1WI datasets, the highest Dice coefficient was observed in the T1_Norm dataset using the DeepLabV3+ model.

Specifically, in the T1_Norm dataset, DeepLabV3+ achieved the highest Dice coefficient of 0.779, while the second-best model was U-Net with a Dice coefficient of 0.709. To evaluate whether the accuracy of DeepLabV3+ was statistically superior to that of U-Net, a paired t-test was conducted. With a significance level (α) set at 0.05, the paired t-test yielded a t-value of 3.4759 and a p-value of 6.9414 × 10⁻⁴. Since p < α, the difference between the two models is statistically significant. Therefore, the DeepLabV3+ model’s accuracy is statistically better than that of the U-Net model.

We also examined the effect of preprocessing on the T1-weighted images and found that the greatest improvement in Dice coefficient occurred when intensity normalization was applied in combination with the DeepLabV3+ model, resulting in a performance gain of 0.156.

In addition, during the training of the DeepLabV3+ model on the T1_Norm dataset, we observed the training and validation loss curves, as shown in Figure 6. Starting from epoch 15, the training loss continued to decrease, while the validation loss plateaued. The increasing gap between the two curves indicates that the model was becoming increasingly specialized in fitting the training data, with limited improvement on the validation data. This suggests that the model showed signs of slight overfitting.

4.4.2. Comparison Results of Model Performance for FLAIR Sequence

The four semantic segmentation models were tested on five different FLAIR datasets. The results are presented in

Table 2. The test results on FLAIR images.

Dataset	Model	Dice	IoU	Recall	Precision
FLAIR_OG	U-Net	0.717	0.559	0.731	0.704
	U-Net++	0.571	0.400	0.586	0.557
	DeepLabV3+	0.631	0.461	0.663	0.602
	LinkNet	0.463	0.301	0.446	0.481
FLAIR_AHE	U-Net	0.739	0.586	0.760	0.719
	U-Net++	0.572	0.401	0.551	0.595
	DeepLabV3+	0.688	0.524	0.679	0.697
	LinkNet	0.451	0.291	0.494	0.415
FLAIR_Norm	U-Net	0.781	0.641	0.764	0.799
	U-Net++	0.626	0.456	0.658	0.597
	DeepLabV3+	0.780	0.639	0.759	0.802
	LinkNet	0.507	0.340	0.523	0.492
FLAIR_BFC	U-Net	0.725	0.569	0.739	0.712
	U-Net++	0.602	0.431	0.635	0.572
	DeepLabV3+	0.682	0.517	0.727	0.642
	LinkNet	0.472	0.309	0.503	0.445
FLAIR_GRAR	U-Net	0.801	0.668	0.826	0.777
	U-Net++	0.640	0.471	0.669	0.613
	DeepLabV3+	0.724	0.567	0.709	0.74
	LinkNet	0.486	0.321	0.495	0.477

.

The four semantic segmentation models were also tested on five different FLAIR image datasets. The results are shown in Table 2. According to the results, the U-Net model consistently achieved the highest Dice coefficient across all datasets. Specifically, it reached 0.717 on the FLAIR_OG dataset, 0.739 on the FLAIR_AHE dataset, 0.781 on the FLAIR_Norm dataset, 0.725 on the FLAIR_BFC dataset, and 0.801 on the FLAIR_GRAR dataset—the highest among all.

We also examined the impact of preprocessing on the FLAIR images and found that the greatest improvement in Dice coefficient was observed when intensity normalization was combined with the DeepLabV3+ model, resulting in a performance gain of 0.149.

In the FLAIR_GRAR dataset, U-Net achieved the highest Dice coefficient of 0.801, followed by DeepLabV3+ with a Dice coefficient of 0.724. A paired t-test was conducted to verify whether the accuracy of the U-Net model was statistically superior to that of DeepLabV3+. With the significance level α set at 0.05, the t-test produced a t-value of 6.0327 and a p-value of 1.5686 × 10⁻⁸. Since p < α, the difference between the two models is statistically significant, indicating that the U-Net model is statistically more accurate than the DeepLabV3+ model.

Additionally, during training with the FLAIR_GRAR dataset, the U-Net model’s training and validation loss curves were observed, as shown in Figure 7. The training loss steadily and consistently decreased, eventually falling below 0.1, indicating that the model learned well on the training data. The validation loss dropped sharply in the early epochs, reaching close to the training loss around epochs 5 to 10. However, from epoch 10 onward, the validation loss began to fluctuate and showed limited further improvement, eventually stabilizing at around 0.17. This behavior suggests that the model exhibited slight signs of overfitting.

4.4.3. Discussion on MRI Semantic Segmentation

From the above experimental results, we found that in the T1WI image sequence, the model with the highest accuracy was DeepLabV3+ combined with intensity normalization, achieving a Dice coefficient of 0.779. The average performance of the DeepLabV3+ model on T1WI images also outperformed other methods, indicating that DeepLabV3+ is a more effective model when using T1WI sequences for training. The semantic segmentation result of one T1WI image sequence using the DeepLabV3+ model is shown in Figure 8.

In the FLAIR image sequence, the highest accuracy was achieved by the U-Net model using the GRAR technique, with a Dice coefficient of 0.801. Similarly, the average performance of the U-Net model with FLAIR images was better than other methods, suggesting that U-Net is the more suitable model for training when FLAIR sequences are used. The semantic segmentation result of one FLAIR image sequence using the U-Net model is shown in Figure 9.

T1WI sequences provide high-resolution anatomical details, offering clear delineation of brain structures. DeepLabV3+ leverages atrous convolution and a robust encoder–decoder architecture, enabling it to capture multi-scale contextual information effectively. This makes it particularly adept at segmenting structures with well-defined boundaries, as commonly seen in T1WI images. Studies have shown that DeepLabV3+ achieves high accuracy and Dice coefficient when applied to T1WI sequences, highlighting its suitability for tasks requiring precise structural segmentation.

FLAIR sequences are designed to suppress cerebrospinal fluid signals, enhancing the visibility of lesions such as white matter hyperintensities. U-Net’s architecture, characterized by its symmetric encoder–decoder structure and skip connections, allows for the preservation of spatial information and fine details. This makes it highly effective in detecting and segmenting lesions that appear prominently in FLAIR images. Research indicates that U-Net models achieve superior segmentation performance on FLAIR sequences.

Moreover, when comparing the various image preprocessing techniques, we observed that using GRAR and intensity normalization as preprocessing methods led to performance improvements across all models. In previous literature, Ullah et al. [12] also demonstrated through experiments that GRAR is an important preprocessing technique that can enhance brain MRI segmentation performance.

When dealing with the complex task of brain tumor medical image segmentation, each model exhibits certain limitations. U-Net relies on transposed convolutions and skip connections for upsampling, but it may fail to capture sufficient detail in cases where tumor boundaries are indistinct or blurry. U-Net++ enhances representational capacity through nested skip connections across different resolutions; however, this complexity increases the risk of overfitting, especially when the medical imaging dataset is small. DeepLabV3+ incorporates atrous convolutions and an ASPP module to improve multi-scale feature extraction, but its performance remains unstable when segmenting very small or poorly defined tumor regions. LinkNet, lacking a multi-scale context module like ASPP, performs poorly in scenarios involving blurry tumor boundaries or multiple small tumor regions.

4.5. Survival Classification Results

In this study, nine survival classification models (DT, RF, XGBoost, SVM, MLP, KNN, GNB, MNB, and BNB) were trained and tested using genetic data. The hyperparameters were set as follows: DT used criterion = “gini”; RF used estimators = 100; XGBoost used estimators = 100; learning rate = 0.3; SVM used a linear kernel (kernel = “linear”); MLP had hidden layer sizes of (16, 5) with solver = “Adam” and activation = “logistic”; KNN used neighbors = 3; MNB used alpha = 1.0; and BNB used binarize = 3 and alpha = 1.0. The dataset was split into 70% for training and 30% for testing.

The results of survival models built using various classification algorithms, after replacing missing values with zero, are shown in

Table 3. The results of survival models for replacing missing values with zero.

Model	Accuracy	Precision	Recall	F1
DT	0.6842	0.9231	0.7059	0.8000
RF	0.7368	0.8750	0.8235	0.8495
XGBoost	0.7895	0.9357	0.8333	0.8824
SVM	0.8421	0.8421	0.8635	0.9143
MLP	0.7895	0.8421	0.8635	0.9143
KNN	0.7692	0.9133	0.9091	0.8696
GNB	0.7368	0.7768	0.9333	0.8485
MNB	0.8421	0.8889	0.9412	0.9143
BNB	0.8947	0.8947	0.9021	0.9444

. Among them, the Bernoulli Bayesian model achieved the highest accuracy and F1-score, reaching 0.8947 and 0.9444, respectively. Based on these two predictive metrics, it can be concluded that this model demonstrates the highest accuracy and relatively superior performance in predicting patient survival.

Table 4. The results of survival models for replacing missing values with one.

Model	Accuracy	Precision	Recall	F1
DT	0.7368	0.9125	0.8667	0.8367
RF	0.7895	0.8833	0.9375	0.8824
XGBoost	0.7868	0.9235	0.8735	0.8485
SVM	0.8421	0.8421	0.8635	0.9143
MLP	0.7895	0.8595	0.8333	0.8824
KNN	0.8462	0.9062	0.9375	0.9167
GNB	0.7368	0.9333	0.7778	0.8485
MNB	0.9474	0.9474	0.9375	0.9730
BNB	0.9474	0.9474	0.9375	0.9730

presents the results of survival models built using various classification algorithms based on patient data where missing values were replaced with 1. It can be observed that both the Bernoulli Bayesian model and the multinomial Bayesian model achieved the highest Accuracy and F1-Score, reaching 0.9474 and 0.9730, respectively. Based on these two predictive indicators, it can be concluded that these models are the most accurate and relatively superior in predicting patient survival.

Table 5. The results of survival models for replacing missing values with two.

Model	Accuracy	Precision	Recall	F1
DT	0.6316	0.7857	0.7333	0.7586
RF	0.7368	0.8750	0.8235	0.8485
XGBoost	0.7868	0.8667	0.8125	0.8387
SVM	0.7895	0.7895	0.8525	0.8824
MLP	0.8421	0.8421	0.8133	0.9143
KNN	0.9231	0.9273	0.9174	0.9600
GNB	0.7268	0.8635	0.8750	0.8585
MNB	0.8421	0.8521	0.8375	0.9143
BNB	0.7885	0.8274	0.8175	0.8824

presents the results of survival models built using various classification algorithms based on patient data where missing values were replaced with 2. It can be observed that the KNN model achieved the highest accuracy and F1-Score, reaching 0.9231 and 0.9600, respectively. Based on these two predictive indicators, it can be concluded that this model demonstrates the highest accuracy and relatively superior performance in predicting patient survival.

From the results above, it can be observed that replacing missing values with different values in genetic data has a significant impact on the analysis outcomes. When the missing values are replaced with 0 or 1, BNB achieves the best performance. However, when the missing values are replaced with 2, representing “unknown”, KNN yields the best results.

5. Conclusions

This study demonstrates that the choice of preprocessing methods significantly influences the performance of deep learning models in brain tumor segmentation. Specifically, the DeepLabV3+ model achieved the best segmentation accuracy on T1WI when combined with intensity normalization, while the U-Net model performed best on FLAIR images with GRAR. These findings highlight the importance of selecting appropriate combinations of models and preprocessing techniques to optimize performance in medical image segmentation tasks.

Preprocessing techniques such as GRAR and intensity normalization consistently improved segmentation outcomes across multiple models, including U-Net, U-Net++, DeepLabV3+, and LinkNet. However, the limited sample size in this study posed a challenge, increasing the risk of overfitting and reducing the model’s generalization ability. Future work can address this limitation through data augmentation or by leveraging Generative Adversarial Networks to expand the training data.

Moreover, traditional evaluation metrics like Dice coefficient may not fully capture the accuracy of tumor boundary and shape delineation. Incorporating additional metrics such as Hausdorff distance and average surface distance could offer a more comprehensive assessment of segmentation quality.

In the genetic data analysis, handling missing values remains a critical issue. Instead of simple binary imputation (e.g., 0 or 1), introducing a third category to represent “unknown” may improve classification accuracy. Additionally, integrating biological pathway analysis tools could provide deeper insights and enhance the biological relevance of survival predictions.

One key limitation of this study is that MRI images and genetic data were not obtained from the same patients, preventing the development of a multimodal model. Future research will focus on collecting both imaging and genetic data from individual patients to enable the construction of integrated models that can more accurately predict survival outcomes through multimodal learning.