Effect of Attention Mechanisms in a Fuzzified Nested U-Net for Medical Image Segmentation

Accurate segmentation of cardiac structures (right ventricle (RV), left ventricle (LV), and myocardium (Myo)) from cardiac MRI images plays an important role in the diagnosis and treatment of cardiovascular disease. However, despite all this, segmentation of these structures at the micro level remains a major challenge due to the complex anatomical diversity and noise inherent in MRI data. This paper examines the performance of various cardiac MRI segmentation techniques, with a primary focus on the overlapping U-Net structure, attention mechanisms, and fuzzy pooling strategies. This study comprehensively evaluates these methods, both independently and in combination, to determine their effectiveness in improving segmentation quality for the RV, LV, and myocardial regions. Additionally, the effect of thresholding strategies on segmentation accuracy is examined. The experimental results on the Automated Cardiac Diagnosis Challenge (ACDC) dataset show that the proposed model (combining nested U-Net, attention mechanisms, and fuzzy pooling) achieved a dice score of 98.20%, an accuracy of 96.83%, and a recall of 96.83%, superior to other basic methods. In comparison, the best-performing core model, ANU-Net, achieved a Dice score of 94.31%, accuracy of 95.19%, and recall of 93.44%. These findings underscore the superior performance of the hybrid model in terms of segmentation and boundary delineation accuracy. These results confirm the potential of hybrid deep learning models in developing cardiac image analysis. Future work will focus on improving these configurations across diverse datasets and also exploring real-time deployment strategies in clinical settings.