This study explores advanced methodologies for enhancing brain tumor segmentation, addressing the complexity and diversity of tumor sub-regions in medical imaging. We introduce a novel approach utilizing Graph Neural Networks (GNNs) that incorporate both spectral and spatial insights for segmentation. By leveraging various supervoxel creation methods such as VCCS, SLIC, Watershed, Meanshift, and Felzenszwalb-Huttenlocher, we structured 3D MRI images into a graph format. This format enabled the implementation of Spectral and Spatial GNNs to capture comprehensive local and global tumor characteristics effectively. Our Spectral-Spatial GNN model, integrating the Laplacian matrix, demonstrated significant improvements in segmenting distinct tumor sub-regions of Necrosis, Edema, and Enhancing Tumor. Our research shows that GNNs, combining spectral and spatial aspects, provide a superior approach for accurate and precise brain tumor segmentation. The preliminary supervoxel generation phase showed VCCS outperformed other methods in homogeneity, inertia, shape uniformity, and size uniformity, setting a strong foundation for further analysis. Hyperparameter tuning maximized the performance of our Spectral- Spatial GNN model. This model demonstrates superior accuracy in both whole tumor and sub- region segmentation compared to traditional methods and individual Spectral or Spatial GNNs, with the Dice coefficient and other metrics as proof. Our Spectral-Spatial GNN model’s accuracy and precision advances have significant potential to enhance brain tumor diagnosis, treatment planning, and fundamental research. In a separate study we introduce a novel hybrid loss function that merges the edge-preservation capabilities of the Mumford-Shah (MS) functional with the precise overlap analysis of the Dice loss, aiming to overcome limitations observed in traditional brain tumor segmentation methods. This approach significantly enhances the accuracy of delineating tumor boundaries and sub-regions. This method advances not only brain tumor segmentation but also presents potential for application in other diagnostic areas where precise segmentation is essential for effective treatment planning. We have observed that finely tuning the balance between the MS and Dice components is vital for optimal performance, emphasizing the need for tailored adjustments in our hybrid loss function for specialized applications. The success of this approach stems from its dual capacity to enhance edge detection and accurately measure overlap, leading to a more detailed and reliable medical image analysis. While primarily focused on brain tumor segmentation, the principles behind our hybrid loss function hold potential for adaptation across various medical imaging tasks, suggesting a broad applicability that could lead to improved diagnostic and therapeutic outcomes.
","abstract_html":"<p>This study explores advanced methodologies for enhancing brain tumor segmentation, addressing the complexity and diversity of tumor sub-regions in medical imaging. We introduce a novel approach utilizing Graph Neural Networks (GNNs) that incorporate both spectral and spatial insights for segmentation. By leveraging various supervoxel creation methods such as VCCS, SLIC, Watershed, Meanshift, and Felzenszwalb-Huttenlocher, we structured 3D MRI images into a graph format. This format enabled the implementation of Spectral and Spatial GNNs to capture comprehensive local and global tumor characteristics effectively. Our Spectral-Spatial GNN model, integrating the Laplacian matrix, demonstrated significant improvements in segmenting distinct tumor sub-regions of Necrosis, Edema, and Enhancing Tumor. Our research shows that GNNs, combining spectral and spatial aspects, provide a superior approach for accurate and precise brain tumor segmentation. The preliminary supervoxel generation phase showed VCCS outperformed other methods in homogeneity, inertia, shape uniformity, and size uniformity, setting a strong foundation for further analysis. Hyperparameter tuning maximized the performance of our Spectral- Spatial GNN model. This model demonstrates superior accuracy in both whole tumor and sub- region segmentation compared to traditional methods and individual Spectral or Spatial GNNs, with the Dice coefficient and other metrics as proof. Our Spectral-Spatial GNN model’s accuracy and precision advances have significant potential to enhance brain tumor diagnosis, treatment planning, and fundamental research. In a separate study we introduce a novel hybrid loss function that merges the edge-preservation capabilities of the Mumford-Shah (MS) functional with the precise overlap analysis of the Dice loss, aiming to overcome limitations observed in traditional brain tumor segmentation methods. This approach significantly enhances the accuracy of delineating tumor boundaries and sub-regions. This method advances not only brain tumor segmentation but also presents potential for application in other diagnostic areas where precise segmentation is essential for effective treatment planning. We have observed that finely tuning the balance between the MS and Dice components is vital for optimal performance, emphasizing the need for tailored adjustments in our hybrid loss function for specialized applications. The success of this approach stems from its dual capacity to enhance edge detection and accurately measure overlap, leading to a more detailed and reliable medical image analysis. While primarily focused on brain tumor segmentation, the principles behind our hybrid loss function hold potential for adaptation across various medical imaging tasks, suggesting a broad applicability that could lead to improved diagnostic and therapeutic outcomes.</p>","abstract_has_math":false,"creators":["Mohammadi, Sina"],"institution":null,"degree_name":"Doctor of Philosophy (PhD)","degree_level":"Dissertation","degree_discipline":"Computational and Data Sciences","degree_department":null,"school":null,"contributors":["Mohamed Allali","Cyril Rakovski","Ahmed Sebbar"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-08-01T07:00:00Z","date_published":"2024-08-01T07:00:00Z","updated_at":"2026-07-24T01:38:37Z","subjects":["Graph Neural Network; Spectral GNN; supervoxel; segmentation","Computer Sciences","Data Science"],"languages":[],"rights":[],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"https://digitalcommons.chapman.edu/cads_dissertations/46","outbound_label":"Repository record","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Mohamed Allali","Cyril Rakovski","Ahmed Sebbar"]},{"key":"dc:creator","label":"Author","values":["Mohammadi, Sina"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.available","label":"Dc Date Available","values":["2026-08-31T07:00:00Z"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Computational and Data Sciences"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Dissertation"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Doctor of Philosophy (PhD)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Graph Neural Network; Spectral GNN; supervoxel; segmentation","Computer Sciences","Data Science"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["https://digitalcommons.chapman.edu/cads_dissertations/46"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["This study explores advanced methodologies for enhancing brain tumor segmentation, addressing the complexity and diversity of tumor sub-regions in medical imaging. We introduce a novel approach utilizing Graph Neural Networks (GNNs) that incorporate both spectral and spatial insights for segmentation. By leveraging various supervoxel creation methods such as VCCS, SLIC, Watershed, Meanshift, and Felzenszwalb-Huttenlocher, we structured 3D MRI images into a graph format. This format enabled the implementation of Spectral and Spatial GNNs to capture comprehensive local and global tumor characteristics effectively. Our Spectral-Spatial GNN model, integrating the Laplacian matrix, demonstrated significant improvements in segmenting distinct tumor sub-regions of Necrosis, Edema, and Enhancing Tumor. Our research shows that GNNs, combining spectral and spatial aspects, provide a superior approach for accurate and precise brain tumor segmentation. The preliminary supervoxel generation phase showed VCCS outperformed other methods in homogeneity, inertia, shape uniformity, and size uniformity, setting a strong foundation for further analysis. Hyperparameter tuning maximized the performance of our Spectral- Spatial GNN model. This model demonstrates superior accuracy in both whole tumor and sub- region segmentation compared to traditional methods and individual Spectral or Spatial GNNs, with the Dice coefficient and other metrics as proof. Our Spectral-Spatial GNN model’s accuracy and precision advances have significant potential to enhance brain tumor diagnosis, treatment planning, and fundamental research. In a separate study we introduce a novel hybrid loss function that merges the edge-preservation capabilities of the Mumford-Shah (MS) functional with the precise overlap analysis of the Dice loss, aiming to overcome limitations observed in traditional brain tumor segmentation methods. This approach significantly enhances the accuracy of delineating tumor boundaries and sub-regions. This method advances not only brain tumor segmentation but also presents potential for application in other diagnostic areas where precise segmentation is essential for effective treatment planning. We have observed that finely tuning the balance between the MS and Dice components is vital for optimal performance, emphasizing the need for tailored adjustments in our hybrid loss function for specialized applications. The success of this approach stems from its dual capacity to enhance edge detection and accurately measure overlap, leading to a more detailed and reliable medical image analysis. While primarily focused on brain tumor segmentation, the principles behind our hybrid loss function hold potential for adaptation across various medical imaging tasks, suggesting a broad applicability that could lead to improved diagnostic and therapeutic outcomes.
"]},{"key":"dc:source","label":"Dc Source","values":["S. Mohammadi, \"Medical image analysis based on graph machine learning and variational methods,\" Ph.D. dissertation, Chapman University, Orange, CA, 2024. https://doi.org/10.36837/chapman.000601"]},{"key":"dc:title","label":"Title","values":["Medical Image Analysis Based on Graph Machine Learning and Variational Methods"]}]}],"canonical_facts":{"dc:contributor":["Mohamed Allali","Cyril Rakovski","Ahmed Sebbar"],"dc:creator":["Mohammadi, Sina"],"dc:date.available":["2026-08-31T07:00:00Z"],"dc:description.abstract":["This study explores advanced methodologies for enhancing brain tumor segmentation, addressing the complexity and diversity of tumor sub-regions in medical imaging. We introduce a novel approach utilizing Graph Neural Networks (GNNs) that incorporate both spectral and spatial insights for segmentation. By leveraging various supervoxel creation methods such as VCCS, SLIC, Watershed, Meanshift, and Felzenszwalb-Huttenlocher, we structured 3D MRI images into a graph format. This format enabled the implementation of Spectral and Spatial GNNs to capture comprehensive local and global tumor characteristics effectively. Our Spectral-Spatial GNN model, integrating the Laplacian matrix, demonstrated significant improvements in segmenting distinct tumor sub-regions of Necrosis, Edema, and Enhancing Tumor. Our research shows that GNNs, combining spectral and spatial aspects, provide a superior approach for accurate and precise brain tumor segmentation. The preliminary supervoxel generation phase showed VCCS outperformed other methods in homogeneity, inertia, shape uniformity, and size uniformity, setting a strong foundation for further analysis. Hyperparameter tuning maximized the performance of our Spectral- Spatial GNN model. This model demonstrates superior accuracy in both whole tumor and sub- region segmentation compared to traditional methods and individual Spectral or Spatial GNNs, with the Dice coefficient and other metrics as proof. Our Spectral-Spatial GNN model’s accuracy and precision advances have significant potential to enhance brain tumor diagnosis, treatment planning, and fundamental research. In a separate study we introduce a novel hybrid loss function that merges the edge-preservation capabilities of the Mumford-Shah (MS) functional with the precise overlap analysis of the Dice loss, aiming to overcome limitations observed in traditional brain tumor segmentation methods. This approach significantly enhances the accuracy of delineating tumor boundaries and sub-regions. This method advances not only brain tumor segmentation but also presents potential for application in other diagnostic areas where precise segmentation is essential for effective treatment planning. We have observed that finely tuning the balance between the MS and Dice components is vital for optimal performance, emphasizing the need for tailored adjustments in our hybrid loss function for specialized applications. The success of this approach stems from its dual capacity to enhance edge detection and accurately measure overlap, leading to a more detailed and reliable medical image analysis. While primarily focused on brain tumor segmentation, the principles behind our hybrid loss function hold potential for adaptation across various medical imaging tasks, suggesting a broad applicability that could lead to improved diagnostic and therapeutic outcomes.
"],"dc:identifier":["https://digitalcommons.chapman.edu/cads_dissertations/46"],"dc:source":["S. Mohammadi, \"Medical image analysis based on graph machine learning and variational methods,\" Ph.D. dissertation, Chapman University, Orange, CA, 2024. https://doi.org/10.36837/chapman.000601"],"dc:subject":["Graph Neural Network; Spectral GNN; supervoxel; segmentation","Computer Sciences","Data Science"],"dc:title":["Medical Image Analysis Based on Graph Machine Learning and Variational Methods"],"thesis:degree_discipline":["Computational and Data Sciences"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-24T01:38:37Z"}