Optimized Attention-Based U-Net with Adaptive Hybrid Loss for Accurate Brain Tumor Segmentation in MRI

Abstract

Brain tumors such as gliomas are a major concern in health care. In order to detect them effectively, practitioners inspect medical images, and MRI is one of the main tools used to analyze brain structures and detect any abnormalities. In this work, we introduce an optimized attention-based deep learning approach for glioma segmentation. Our contributions include : (1) optimized attention gates that use Hard Sigmoid activation and 1×1 convolutions to reduce computational complexity from O(C^2) to O(C) while preserving the discriminative capacity of the original gates. (2) An adaptive loss weighting mechanism that adjusts the balance between Dice Loss and Binary Cross-Entropy during training in order to improve learning stability, and (3) a joint segmentation-classification framework that leverages global constraints to reduce False Positives. We have validated this approach on BRATS 2018 and Lower Grade Glioma (LGG) datasets. On BRATS, we achieve a 94.84% Dice coefficient, 90.16% IoU and 96.01% precision; on LGG, 92.85% Dice, 87.15% IoU and 89.76% precision. This corresponds to the best precision on both datasets among all compared methods. We have also noted that the optimal value of the adaptive weight of the loss function, αmax = 0.7, maximizes the Intersection over Union on BRATS, and that the classification branch under this configuration produces substantially fewer false positives compared to the ablation where classification is not used.