Evaluation of EMA-Attention U-Net for Brain Hemorrhage Segmentation

Abstract

This paper applies an enhanced multi-class segmentation approach for brain hemorrhage detection on head CT images. We address the limitation of standard architectures that struggle to delineate lesion boundaries under complex multi-class conditions, where variations in size, shape, and visual similarity are significant. We utilize an improved Attention U-Net by replacing the conventional attention mechanism with Efficient Multi-Scale Attention (EMA), aiming to strengthen feature representation while preserving the original encoder–decoder structure. We train and evaluate both the baseline and the proposed models on a head CT dataset with COCO-based annotations converted into pixel-wise masks, using metrics including Dice coefficient, Intersection over Union (IoU), precision, recall, parameter count, and inference speed. This study obtains clear performance improvements from the proposed EMA-Attention U-Net. We achieve a higher validation Dice score of 0.7836 compared to 0.7396 from the baseline. On the global macro test evaluation, we observe significant gains in Dice (0.1420 to 0.2898), IoU (0.0982 to 0.2475), and precision (0.2031 to 0.3837). Although recall slightly decreases from 0.6927 to 0.6476 and inference speed reduces from 133.12 FPS to 91.15 FPS, we maintain a nearly identical parameter count (7,791,285 to 7,794,805). These results show that we can improve spatial accuracy and segmentation consistency with minimal computational overhead, confirming that EMA is an effective attention mechanism for complex multi-class medical image segmentation.