Brain Image Segmentation: Advancements, Applications, and Challenges in Neuroimaging

Brain image segmentation is a cornerstone of modern medical image analysis, serving as the initial and often most critical step in numerous clinical and research applications. This process involves partitioning an input image, typically derived from Magnetic Resonance Imaging (MRI), into distinct regions where pixels or voxels share similar attributes. This fundamental operation allows for the precise isolation and analysis of specific anatomical structures, such as white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF), as well as pathological areas like tumors or lesions.¹ By transforming complex raw imaging data into a more streamlined and interpretable representation, segmentation significantly enhances the utility of these images for diagnosis, treatment planning, and scientific inquiry. A key principle for effective brain MRI segmentation lies in leveraging spatial context; the intensity of a given pixel or voxel is statistically dependent on its neighbors, a characteristic that is exploited by various algorithms to achieve coherent region delineation.⁴ The accuracy of this initial segmentation step profoundly influences the outcome of subsequent analyses, underscoring its foundational importance in the entire neuroimaging pipeline.⁴

Clinical and Research Applications of Brain Image Segmentation

In clinical settings, brain MRI segmentation is indispensable for several critical applications. It is routinely employed for measuring and visualizing the brain’s anatomical structures, providing quantitative data essential for monitoring brain changes over time, such as atrophy in neurodegenerative diseases.⁴ Furthermore, it plays a vital role in delineating pathological regions, including lesions and tumors, enabling clinicians to accurately understand a patient’s condition.⁵ This precise differentiation between tumor and normal tissues is crucial for ascertaining tumor type and subsequently informing critical decisions regarding surgical margins, radiation therapy, and chemotherapy.⁷

The advent of intelligent solutions integrating deep learning with computer vision has markedly increased the level of automation in image analysis. This technological advancement significantly enhances diagnostic efficiency and objectivity by automating the segmentation of tumor regions, thereby reducing the workload on physicians and facilitating the precise formulation of treatment plans, leading to more timely and effective patient care.⁷ The high precision demanded in clinical practice for brain tumor segmentation is paramount to minimize misdiagnoses and oversights, optimize treatment strategies, and ultimately enhance patient outcomes.⁷

From a research perspective, brain image segmentation opens new horizons for studying brain anatomy and function. The rapid development of non-invasive brain imaging technologies, particularly MRI, has provided vast amounts of high-quality data, making segmentation an essential tool for analyzing these complex datasets.⁴

Segmentation is crucial for understanding tumor biology, predicting prognosis, and estimating survival for brain malignancies.⁹ Researchers leverage multi-modal MRI images, which include T1-weighted (T1), T2-weighted (T2), T1-weighted with contrast enhancement (T1ce), and fluid attenuation inversion recovery (FLAIR) modalities, to obtain a comprehensive and accurate depiction of tumors and adjacent brain tissue.⁹ The integration of these modalities provides additional information vital for diagnosing and monitoring brain tumors, allowing researchers to assess the efficacy of various segmentation algorithms and compare their outcomes. This comparative analysis is instrumental in stimulating the development of novel methodologies and improving the precision of brain tumor segmentation.

The increasing prevalence of brain disorders has driven a significant demand for automated brain segmentation solutions. These automated methods assist medical experts in early diagnosis and treatment interventions, highlighting the critical role of segmentation in addressing public health challenges.¹³ The shift from time-consuming and variable manual segmentation to automated solutions is a direct response to the need for consistent and accurate delineation of target brain regions and tumors for clinical evaluation of disease progression, surgical planning, post-surgical matching, and radiation therapy outcomes.¹³

Traditional Brain Image Segmentation Methods

Historically, brain image segmentation relied on a suite of traditional techniques that laid the groundwork for more advanced approaches. These methods are generally classified based on the image information they employ.¹⁷

Overview of Traditional Techniques

Traditional image segmentation techniques, while often simpler conceptually, were foundational in the field before the widespread adoption of machine learning and deep learning. These methods include thresholding, region-based segmentation, edge-based segmentation, and clustering-based segmentation. Manual segmentation, though a human-driven process, is also considered a traditional approach due to its historical prevalence and its role as a “ground truth” benchmark for automated methods.⁴

Detailed Explanation of Key Traditional Methods

Thresholding

Thresholding is one of the simplest and fastest segmentation approaches, particularly effective for two-dimensional images where the difference in pixel values between target classes is high. This method converts an image into a binary map by assigning pixels above a certain intensity threshold to one class (e.g., 1) and those below to another (e.g., 0), a process often termed binarization. While conceptually straightforward and useful for tasks like contour detection, its primary limitation lies in considering only the intensity value of the current pixel, neglecting the relationships between neighboring pixels.¹⁷ This oversight means threshold-based methods often fail to fully exploit the rich information provided by MRI slices and are typically used to separate and eliminate the background rather than delineate complex brain structures.¹⁷

Region-Based Segmentation

Region-based segmentation approaches partition an image into regions based on predefined similarity criteria, often starting from a “seed” pixel or group of pixels. The algorithm then expands by checking neighbors of the seeds and adding pixels that satisfy the similarity criteria (e.g., based on intensity or edges) to the growing region. A key characteristic of region growing is its ability to segment similar regions and generate related areas.¹⁷ However, these methods are sensitive to noise, which can lead to holes in extracted regions.¹⁷ A significant disadvantage is the Partial Volume (PV) effect, where a single voxel may contain multiple tissue types, blurring borders and limiting segmentation accuracy.¹⁷ Additionally, the choice of the initial seed point is critical; an improperly chosen seed can cause the region to grow outside the object of interest or merge with an unintended region.¹⁷

Edge-Based Segmentation

Edge-based segmentation, also known as edge detection, focuses on identifying sharp changes in image intensity that typically correspond to object boundaries. From a segmentation perspective, this involves classifying pixels as either “edge pixels” or “non-edge pixels.” This is commonly achieved using special filters that estimate image gradients in the spatial plane, such as the Canny edge detection algorithm. Edge detection is valuable for stripping images of redundant information, thereby reducing their size and facilitating further analysis.

Clustering-Based Segmentation

Clustering algorithms represent a more modern traditional approach, often performing better than simpler counterparts by grouping pixels with common attributes into “k” classes without requiring initial seed points. Popular algorithms like K-means clustering are unsupervised, meaning they learn patterns directly from the data. While generally efficient, methods like Fuzzy C-means (FCM), a common clustering technique in medical image segmentation, can produce suboptimal results in the presence of noise and intensity inhomogeneity.¹⁸ This highlights a challenge in applying these methods directly to complex and often noisy medical images.

Manual Segmentation

Manual segmentation, performed by human operators such as expert physicians, is traditionally considered the most accurate method for delineating structures in medical images.⁴ This process typically involves segmenting and labeling images slice-by-slice for 3D volumetric data.⁴ Despite its perceived accuracy, manual segmentation is an intensive and time-consuming task, requiring a trained operator to process numerous slices for each target structure.⁴ It is also particularly prone to errors, as evidenced by intra- or inter-operator variability studies, and results are often difficult or impossible to reproduce even by the same experienced operator. Despite these significant drawbacks, manual segmentation remains intensively used for defining “ground truth” data, which serves as a surrogate for true delineation and is essential for quantitatively evaluating automated segmentation methods.⁴ It is also fundamental for brain atlas formation, which is then used in atlas-based segmentation approaches.⁴ The inherent difficulties and inconsistencies of manual methods underscore the critical need for robust automated solutions.

Modern Brain Image Segmentation Approaches: Machine Learning and Deep Learning

Deep learning, a subset of machine learning, employs multi-layered neural networks, known as deep neural networks, to emulate the complex decision-making processes of the human brain. This paradigm shift has enabled intelligent solutions that integrate deep learning with computer vision, significantly increasing the level of automation in image analysis and enhancing diagnostic efficiency and objectivity.⁷

Convolutional Neural Networks (CNNs)

Convolutional Neural Networks (CNNs or ConvNets) are a cornerstone of deep learning in computer vision, particularly adept at image classification and segmentation tasks. They are designed to automatically extract hierarchical features from images, detecting patterns from simple elements like colors and edges in earlier layers to larger shapes and complex structures in deeper layers. This is achieved through a multi-layered architecture typically comprising convolutional layers, pooling layers, and fully connected layers. CNNs have demonstrated superior performance with image, speech, and audio signal inputs compared to earlier methods, revolutionizing tasks like object detection and image recognition. In medical imaging, CNNs are highly efficient for image segmentation, enhancing the precision of tumor detection and overall diagnostic accuracy by automatically adapting the feature space of the input image. Despite their powerful capabilities, CNNs are computationally demanding, often requiring significant graphical processing unit (GPU) resources, and necessitate highly trained experts for configuration and testing.

U-Net Architecture

The U-Net architecture has emerged as a benchmark model for biomedical image segmentation, particularly in brain tumor segmentation, due to its efficacy in precisely localizing areas of interest. Introduced in 2015, U-Net is a convolutional neural network characterized by its distinctive encoder-decoder structure, which includes “skip connections”. The encoder path (contracting path) captures contextual information by progressively downsampling the image, while the decoder path (expansive path) reconstructs the segmentation map by upsampling.¹⁰ The skip connections concatenate feature maps from the contracting path to the corresponding expansive path, providing higher-resolution features that help recover spatial information lost during downsampling, thereby improving localization accuracy.¹⁰ This design has enabled U-Net to perform exceptionally well, with Dice Similarity Coefficient (DSC) scores often exceeding 0.90 in brain tumor segmentation tasks. Its simple yet effective design has made it a widely used foundational model in medical image segmentation.⁷ Furthermore, enhanced U-Net models, such as Residual U-Net and Ensemble U-Net, have been proposed to further improve performance, addressing limitations like unclear boundaries and over-segmentation.⁷

Vision Transformers (ViTs) and Attention Mechanisms

Vision Transformers (ViTs) represent a powerful evolution in image segmentation, particularly gaining prominence in recent years. Unlike traditional CNNs, ViTs process entire images as sequences of patches, leveraging self-attention mechanisms to capture long-range dependencies and global context across the image. This global understanding is crucial for accurate object detection and segmentation, addressing a limitation of some CNNs in aggregating semantic information across various scales.

Attention mechanisms are a key component, dynamically modulating the weights assigned to different regions within the feature map. This enhancement sharpens the model’s focus on target-related features, ameliorates its capacity to select and integrate features, and bolsters its ability to handle multi-scale targets while enhancing the utilization of contextual information.⁷ An example is the Attention Gate Residual U-Net (AGResUNet), which effectively filters and suppresses irrelevant or noisy feature information while significantly enhancing the expression of features closely related to the tumor.⁷ The ability of ViTs to capture global dependencies makes them highly effective in scenarios where traditional CNNs might struggle with complex features and unclear boundaries.

One-Stage Object Detection (YOLO)

You Only Look Once (YOLO) is an innovative one-stage object detection algorithm that has found application in medical image segmentation, including brain tumors.⁷ This framework utilizes a single neural network as its core architecture, achieving end-to-end object detection tasks directly and accurately predicting coordinates and object positions.⁷ Compared to traditional image segmentation models, YOLO can significantly reduce model complexity and computational costs while maintaining high detection accuracy, demonstrating high versatility and transferability.⁷ This efficiency makes it suitable for real-time clinical work.⁷

Improved YOLO models, such as YOLOv5s-ASPP, integrate Atrous Spatial Pyramid Pooling (ASPP) to enhance multi-scale feature representation, expanding the receptive field and bolstering the model’s capacity to detect targets across diverse scales.⁷ Subsequently, attention mechanisms are intercalated before each detection head to dynamically modulate weights, further sharpening the model’s focus on target-related features and improving the handling of multi-scale targets.⁷

Emerging Techniques and Future Directions

The field of brain image segmentation continues to evolve rapidly, driven by the demand for more accurate, efficient, and adaptable solutions.

Foundation Models and Segment Anything Model (SAM)

A significant emerging trend is the development of medical foundation models, particularly 3D models for multimodal neuroimage segmentation. These models leverage self-supervised training with vision transformers, significantly reducing the extensive data requirements typically necessary for AI model training in neuroimage segmentation. This dual-phase methodology offers the flexibility to adapt to various imaging modalities and addresses the challenge of limited patient data, especially for rare conditions like brain tumors.

The Segment Anything Model (SAM), introduced as part of the Segment Anything project, represents a paradigm shift in image segmentation. SAM harnesses prompt-based learning, enabling more dynamic and adaptable segmentation models that require less manual tuning and effort.²² While current SAM models primarily function in 2D space, requiring 3D modalities to be sub-sampled or processed in slices, ongoing research aims to adapt these models to fully function in 3D, recognizing that the most valuable information is often found in 3D modalities like MRI. This adaptation is expected to enhance the accuracy and efficiency of medical diagnoses and treatment planning by providing precise segmentation.²²

Patch-wise U-Net and Unsupervised Clustering

Innovations in U-Net architectures also include patch-wise U-Net, which processes non-overlapping patches from MRI slices.²³ This approach helps overcome some drawbacks of conventional U-Net by retaining more local information, thereby improving localization accuracy in tissue segmentation.²³

Furthermore, novel approaches are integrating unsupervised clustering with conditional spatial properties. For instance, an intuitionistic-based clustering technique incorporates a nuanced understanding of uncertainty inherent in image data by calculating a hesitation degree.²² This method introduces a conditional spatial function alongside the intuitionistic membership matrix, allowing for the consideration of spatial relationships within the image. Such approaches aim to enhance robustness with homogenous segments, lower sensitivity to noise and intensity inhomogeneity, and accommodate the inherent uncertainty in real-world datasets.²²

Multi-modal Data Integration and Computational Efficiency

The trend towards integrating multi-modal data is becoming increasingly important. Combining different MRI modalities (T1, T2, T1ce, FLAIR) provides a more comprehensive and accurate depiction of tumors and adjacent brain tissue, which is essential for successful segmentation. This multi-modal approach is crucial for improving the precision of brain tumor segmentation and stimulating the development of novel methodologies.

Finally, a continuous focus in the field is to explore more efficient and accurate algorithms that reduce model complexity and computational costs while maintaining high detection accuracy. This pursuit is vital for enabling real-time clinical deployment and making advanced segmentation techniques more accessible in healthcare settings.⁷

Challenges and Limitations in Brain Image Segmentation

Despite significant advancements, brain image segmentation continues to face several inherent challenges and limitations stemming from the nature of medical images and the complexities of the human brain.

Inherent Image Characteristics

Noise and Artifacts

Brain MRI images are susceptible to various forms of noise and artifacts, which complicate the segmentation process and can lead to inaccurate results, particularly for intensity-based segmentation techniques.² These include biological noise (e.g., patient movement) and scanner noise due to tissue non-uniformity and hardware limitations.² A pervasive issue is intensity inhomogeneity, which manifests as a smooth shading effect on reconstructed MR images due to non-uniformity in the radio-frequency coil. This artifact can significantly deteriorate the performance of any image processing algorithm that relies on intensities as a feature. Preprocessing steps are often required to mitigate these issues, but they remain a persistent challenge.²

Partial Volume Effect (PVE)

The Partial Volume Effect (PVE) is arguably one of the most crucial unsolved problems in brain MRI segmentation.¹⁵ It occurs because the finite spatial resolution of imaging devices means that a single image voxel may contain several types of tissues (e.g., both gray matter and white matter), leading to image intensities that do not purely represent any single underlying tissue class.² This phenomenon blurs the borders between different tissue types and significantly complicates the segmentation process, especially given the complex anatomy of the human brain.¹⁷ Ignoring PVE can lead to substantial volume measurement errors, ranging from 20% to 60%.²⁵ Furthermore, resampling images to a stereotactic space, a common preprocessing step, can inadvertently amplify the PVE, potentially leading to incorrect cortical thickness computations if hard tissue classification is used.²⁵

High Similarities and Correlations

Brain segmentation is inherently difficult due to the high similarities and correlations of image intensity among different brain regions.² This makes it challenging for algorithms to precisely distinguish between adjacent structures that may have similar signal characteristics, requiring sophisticated methods that can leverage spatial context effectively.⁴

Methodological and Practical Limitations

Manual Segmentation Drawbacks

While manual segmentation by clinical experts is considered the highest standard for defining “ground truth,” it presents significant practical limitations.¹⁹ The process is extremely time-consuming, tedious, and not practical for labeling large datasets, often requiring a trained operator to review dozens of slices per 3D volume.⁴ Moreover, it is prone to errors and suffers from limitations in inter-rater and intra-rater reproducibility, meaning results can vary significantly between different experts or even for the same expert at different times. This variability compromises diagnostic consistency and accuracy, highlighting the urgent need for automated, reproducible solutions.

Generalizability and Robustness

Ensuring that segmentation models generalize well across diverse patient populations, different MRI scanners, varying acquisition protocols, and a range of image qualities remains a significant challenge. Models trained on specific datasets may not perform robustly on unseen data from different domains. This necessitates advanced techniques like transfer learning and data augmentation to improve model robustness and adaptability.

Ambiguous Boundaries and Over/Under-segmentation

Segmenting complex features with unclear or ambiguous boundaries is difficult for many algorithms, leading to issues such as over-segmentation (where a single region is split into multiple parts) or under-segmentation (where multiple regions are merged).²¹ This is particularly problematic in areas with subtle distinctions between tissue types or in the presence of pathologies with irregular shapes.²¹ Models relying heavily on skip connections, like some U-Net variants, can sometimes suffer from over-segmentation due to redundancy.²¹

Software Tools and Libraries for Brain Image Segmentation

The development of sophisticated software tools and programming libraries has been instrumental in advancing brain image segmentation, making complex algorithms accessible to researchers and clinicians.

Dedicated Software Platforms

Several comprehensive software platforms are widely used for brain image segmentation and neuroimaging analysis:

• BrainSuite: This is a collection of open-source software tools designed for largely automated processing of human brain MRI.²⁸ Its core functionalities include extracting and parameterizing cortical surfaces, segmenting and labeling gray and white matter structures, and analyzing diffusion and functional MRI data.²⁸ BrainSuite provides a user-friendly graphical interface (GUI) available for Windows, Mac OS, and Linux, and integrates atlases like the USCBrain Atlas for structural labeling.²⁸ It also offers a BIDS App for containerized workflows, enhancing reproducibility and ease of deployment.²⁸
• 3D Slicer: A free and open-source image computing platform, 3D Slicer is highly versatile for visualization, processing, segmentation, registration, and analysis of medical and biomedical 3D images. For segmentation, it supports 2D, 3D, and 4D images, allowing for hundreds of segments per image through its Segment Editor module. 3D Slicer incorporates AI-assisted segmentation features, including NVIDIA Clara AI and compatibility with TensorFlow and MONAI, and provides tools for generating ground truth training data. Its extensibility through numerous modules allows for specialized applications like dental image analysis (SlicerCMF) and diffusion MRI (SlicerDMRI).
• FSL (FMRIB Software Library): Developed by the University of Oxford, FSL is a comprehensive library of image analysis and statistical tools for fMRI, MRI, and DTI brain imaging data.³⁰ It includes a dedicated tool for segmentation called FAST (FMRIB’s Automated Segmentation Tool), which allows researchers to identify different tissue types and structures within brain imaging data.³⁰ FSL is widely used in neuroscience research for structural imaging, segmentation, and registering data to standard spaces.³⁰
• FreeSurfer: This free software package, developed at the Martinos Center for Biomedical Imaging, is widely used for creating computerized models of the brain from MRI data.³² FreeSurfer automatically segments most macroscopically visible brain structures, including volumetric segmentation of cortical and subcortical regions, and hippocampal subfields.³² It also measures morphometric properties such as cortical thickness and curvature, and generates topologically correct cortical surface representations.³²

Programming Libraries and Frameworks

For researchers and developers, several programming libraries and frameworks facilitate the implementation of advanced segmentation algorithms:

• TorchIO: An open-source Python library, TorchIO is designed to streamline the loading, preprocessing, augmentation, and patch-based sampling of medical images for deep learning applications.⁹ It integrates seamlessly with PyTorch and is compatible with other frameworks like TensorFlow and MONAI. Key features include spatial and intensity normalization, resampling, reorientation, and the simulation of MRI-specific artifacts (e.g., k-space spike, motion, ghosting, bias field).⁹ TorchIO’s support for patch-based sampling helps overcome memory limitations, and its design promotes experiment reproducibility and traceability.⁹
• MedicalZooPytorch: This PyTorch-based deep learning framework is specifically developed for multi-modal 2D/3D medical image segmentation, with an initial focus on brain MRI.³³ It implements various state-of-the-art 3D deep neural network architectures, including U-Net3D, V-Net, ResNet3D-VAE, and DenseVoxelNet.³³ The library also provides comprehensive data loaders for common medical imaging datasets such as BraTS and Iseg, emphasizing open and reproducible deep learning research.³³

Works cited

1. CNN-Based Image Segmentation Approach in Brain Tumor Classification: A Review – MDPI, accessed on July 21, 2025, https://www.mdpi.com/2673-4591/84/1/66
2. Challenges in Brain Magnetic Resonance Image Segmentation – American Scientific Research Journal for Engineering, Technology, and Sciences, accessed on July 21, 2025, https://asrjetsjournal.org/index.php/American_Scientific_Journal/article/download/2571/1020/0
3. Challenges in Brain Magnetic Resonance Image Segmentation | American Scientific Research Journal for Engineering, Technology, and Sciences, accessed on July 21, 2025, https://asrjetsjournal.org/index.php/American_Scientific_Journal/article/view/2571
4. Automated Whole Brain Segmentation Using FreeSurfer, accessed on July 21, 2025, https://adni.bitbucket.io/reference/docs/UCSFFRESFR/UCSFFreeSurferMethodsSummary.pdf
5. BrainSegFounder: Towards Foundation Models for Neuroimage Segmentation – arXiv, accessed on July 21, 2025, https://arxiv.org/html/2406.10395
7. Application of MRI image segmentation algorithm for brain tumors based on improved YOLO, accessed on July 21, 2025, https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1510175/full
8. Brain Segmentation | Papers With Code, accessed on July 21, 2025, https://paperswithcode.com/task/brain-segmentation
9. Recent deep learning-based brain tumor segmentation models using multi-modality magnetic resonance imaging: a prospective survey – Frontiers, accessed on July 21, 2025, https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2024.1392807/full
10. Recent deep learning-based brain tumor segmentation models using multi-modality magnetic resonance imaging: a prospective survey – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/publication/382446509_Recent_deep_learning-based_brain_tumor_segmentation_models_using_multi-modality_magnetic_resonance_imaging_a_prospective_survey
11. start · BrainSuite, accessed on July 21, 2025, https://brainsuite.org/
12. 3D Slicer image computing platform | 3D Slicer, accessed on July 21, 2025, https://www.slicer.org/
13. Brain Image Segmentation in Recent Years: A Narrative Review – PMC, accessed on July 21, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC8392552/
14. Brain Image Segmentation in Recent Years: A Narrative Review – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/publication/353806079_Brain_Image_Segmentation_in_Recent_Years_A_Narrative_Review
15. Partial Volume Segmentation of Brain MRI Scans of Any Resolution and Contrast, accessed on July 21, 2025, https://www.researchgate.net/publication/346062040_Partial_Volume_Segmentation_of_Brain_MRI_Scans_of_Any_Resolution_and_Contrast
18. Brain Tumour Segmentation Using Choquet Integrals and Coalition Game – MDPI, accessed on July 21, 2025, https://www.mdpi.com/2078-2489/16/7/615
21. UNet limitations in brain tissue segmentation | Download Scientific Diagram – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/figure/UNet-limitations-in-brain-tissue-segmentation_fig2_364330795
22. Intensity Inhomogeneity Correction of Magnetic Resonance Images using Patches – PMC, accessed on July 21, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC4112202/
23. black0017/MedicalZooPytorch: A pytorch-based deep … – GitHub, accessed on July 21, 2025, https://github.com/black0017/MedicalZooPytorch
24. FreeSurfer – PMC, accessed on July 21, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC3685476/
25. Image Segmentation for Intensity Inhomogeneity in Presence of High Noise – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/publication/324616860_Image_Segmentation_for_Intensity_Inhomogeneity_in_Presence_of_High_Noise
26. Partial volume effect modeling for segmentation and tissue classification of brain magnetic resonance images: A review, accessed on July 21, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC4241492/
27. TorchIO: A Python library for efficient loading, preprocessing …, accessed on July 21, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC8542803/
29. Segment Anything – CVF Open Access, accessed on July 21, 2025, https://openaccess.thecvf.com/content/ICCV2023/papers/Kirillov_Segment_Anything_ICCV_2023_paper.pdf
32. Advancing Brain MRI Segmentation using Segment Anything Model – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/publication/390099091_Advancing_Brain_MRI_Segmentation_using_Segment_Anything_Model
33. FSL: Tool/Resource Info – NITRC, accessed on July 21, 2025, https://www.nitrc.org/projects/fsl/
35. (PDF) A Review on Brain MRI Image Segmentation – ResearchGate, accessed on July 21, 2025, https://www.researchgate.net/publication/258859479_A_Review_on_Brain_MRI_Image_Segmentation