Segmentation of Medical MRI Images Using Nested U-Net with Attention Mechanism and Fuzzy Pooling
Noor M. Basheer, Ali Al-Saegh · Maǧallaẗ al-handasaẗ al-rāfidayn · 2025
Segmentation of medical images is a crucial step in the diagnosis and treatment of various diseases, particularly when analyzing magnetic resonance imaging (MRI) scans. Despite significant progress, accurate segmentation remains challenging due to the complexity of anatomical structures, ambiguous boundaries, and variations in image quality. This paper proposes an enhanced Nested U-Net structure that incorporates an attention mechanism and fuzzy pooling to improve retail performance. Nested U-Net benefits from extensive skip connections to capture multi-level contextual information, while the attention mechanism enhances the model's ability to focus on relevant features and noise suppression. In addition, we replace the traditional extreme pooling layers with fuzzy pooling, which enables the network to handle spatial ambiguity more effectively and produce more accurate and refined retail maps. Experimental results on publicly available MRI datasets show that the proposed model achieves remarkable improvements, with a 4.97% increase in the dice coefficient, a 2.76% improvement in accuracy, and a 4.25% increase in recall compared to enhanced U-Net structures. Furthermore, significant improvements have been observed in Hausdorff's distance measures.