BRAIN TUMOR SEGMENTATION USING MULTIATTRIBUTE AGGREGATION AND ADAPTIVE GRADIENT DICE LOSS MECHANISM

Abstract

The human brain is a remarkable entity, central to all bodily functions. Given its critical role, prompt reporting of any irregularities mainly tumors that further lead to cancer in its structure is vital for reducing mortality rates. Segmenting abnormal regions is key for effective treatment planning and monitoring. The most crucial task in this process is the differentiation of healthy tissue from abnormal areas. To date, various imaging techniques have been employed for the early detection of these anomalies, with Magnetic Resonance Imaging (MRI) standing out as a notable and noninvasive diagnostic tool. This research presents an integrated architecture Multi-Attribute Aggregation (MAA) and an Adaptive Gradient Dice Loss (AGDL) (MAA+AGDL) mechanism to significantly improve the accuracy of segmenting brain tumors from MRI scans. The novel framework employs a dual-path architecture that processes foreground and background regions separately, incorporating enhanced feature extraction through improved encoders, a multi-attribute aggregation module for comprehensive feature representation, and mapping prediction modules for precise segmentation. To overcome traditional challenges associated with gradient vanishing and feature misalignment, the framework integrates an AGDL mechanism for stable training and a Deformable Convolution-based Feature Alignment (DCFA) for adaptive feature alignment. Evaluated on the BRATS 2019 and 2020 datasets, our model demonstrates superior performance over existing approaches, achieving higher Dice Similarity Coefficient scores, improved sensitivity, and reduced Hausdorff Distance. These advancements signify a substantial step forward in the use of deep learning for medical imaging, promising enhanced clinical diagnostics and treatment planning for brain tumor patients.