Hybrid deep learning framework for brain tumor type, grade, and segmentation in MRI images

Suliman Farhan, Ahmeed

Abstract

Brain tumours are a critical global health challenge, accounting for 85-90% of all primary central nervous system tumours, with an estimated 308,102 new cases globally in 2020. These tumours often lead to severe physical and cognitive impacts on patients, but early and accurate diagnosis is essential for improving survival rates through better treatment planning. Current diagnostic methods rely heavily on radiologist expertise, making the process time-consuming, expensive, and resource-intensive, placing strain on healthcare systems. In this thesis, we aim to address this diagnostic bottleneck by developing a unified hybrid deep learning framework namely Brain Tumor Diagnostics and Analysis (TDA) that aims to produce accurate diagnostic output for several vital processes within Brain Tumour Diagnostics including classification, grading, and segmentation in MRI images. A novel ensemble dual-modality approach for 3D brain tumor segmentation is proposed which combines dual MRI modalities to enhance segmentation accuracy. The ensemble model, integrating the best-performing dual-modality combinations, achieved a Dice Coefficient of 97.73% and a Mean IoU of 60.08% on the BraTS2020 dataset, significantly outperforming single-modality and dual-modality models. A parallel residual convolutional network (PRCnet) is proposed for tumour classification which leverages advanced techniques such as parallel layers with varied filter sizes and dropout layers to achieve an accuracy of 94.77% on dataset A and 97.1% on dataset B respectively. The framework further addresses key challenges in medical imaging through the development of novel data augmentation method namely orientated Combination MRI (OCMRI) to further enhance performance by mitigating class imbalance and data scarcity. OCMRI, through the controlled fusion of MRI images based on mean squared error thresholds, improve PRCnet classification accuracy. Additionally, the Brain Tumor Diagnostics and Analysis (TDA) framework includes an interactive web-based interface to support clinical interaction, enabling feedback-driven refinement through dynamic validation mechanisms. Also, the inclusion of explainability techniques such as Grad-CAM enhances clinical trust and interpretability. Overall, this thesis contributes an end-to-end, adaptable, and explainable AI-based diagnostic solution for brain tumor analysis, aiming to support radiologists and improve clinical decision-making processes.

Citation

Suliman Farhan, A. (2025). Hybrid deep learning framework for brain tumor type, grade, and segmentation in MRI images. (Thesis). University of Hull. https://hull-repository.worktribe.com/output/5292329

• Thesis Type: Thesis
• Deposit Date: Jul 31, 2025
• Publicly Available Date: Jun 12, 2030
• Keywords: Computer science
• Public URL: https://hull-repository.worktribe.com/output/5292329
• Additional Information: School of Computer Science; Faculty of Engineering and Science; University of Hull
• Award Date: Jun 11, 2025