Investigating diverse loss functions for myocardium ring segmentation in Cardiac Magnetic Resonance images using fuzzy pooling

Abstract

Cardiovascular disease, a leading cause of mortality, underscores the critical need for precise diagnostic methods. Cardiac Magnetic Resonance (CMR) imaging is pivotal for diagnosing heart conditions, yet accurately segmenting the myocardium ring (MYO) remains a significant challenge. This study enhances the U-Net model with fuzzy pooling and evaluates the effects of different loss functions: cross-entropy loss, which evaluates the disparity between predicted and actual probability distributions; focal loss, which tackles class imbalance by reducing the weight of easily classified examples; dice loss, which emphasizes the overlap between predicted and actual segments; Lovász-Softmax loss, which is optimized for Intersection over Union (IoU); and CrossLov, which merges cross-entropy and Lovász-softmax, using the ACDC 2017 dataset. Focal loss achieved the lowest train loss scores of 0.0011% at epoch 95 and 0.0012% at epoch 96. Cross-entropy showed high dice scores but did not excel in boundary delineation. Dice loss showed moderate performance. Lovász-softmax excelled in IoU with an average score of 90.68%, while CrossLov exhibited balanced performance, achieving robust general segmentation results with an IoU score of 93.691%. Additionally, CrossLov attained the lowest Hausdorff Distance (HD), with an overall score of 2.816 mm and 1.309 mm for the MYO, indicating superior boundary precision. These findings highlight the role of loss function selection used together with fuzzy pooling for enhancing the robustness and precision of MYO segmentation, thereby contributing to improved diagnostic accuracy in cardiovascular care.