Generative Adversarial Network-Driven Segmentation of Myocardial Fibrosis in Transthoracic Echocardiography

Abstract

Cardiosclerosis-related myocardial fibrosis is a major cause of impaired cardiac contractility and adverse outcomes, yet its delineation on transthoracic echocardiography (TTE) remains challenging due to low contrast, speckle noise, and weak boundary definition. This study proposes a GAN-driven segmentation framework to automatically detect and segment fibrotic myocardial regions in apical 4-chamber (A4C) TTE images. The method employs a U-Net - based generator to produce binary fibrosis masks and a PatchGAN discriminator to enforce realistic local boundary structure through adversarial learning. A clinically grounded dataset was formed from anonymized echocardiography of 47 confirmed cardiosclerosis patients (Philips EPIQ, GE Vivid), complemented by open datasets (EchoNet-Dynamic, CAMUS) to enhance generalization. Expert cardiologists produced ground-truth masks, with high inter-rater agreement. The model was trained using a combined loss function (adversarial + Dice + binary cross-entropy) with systematic augmentation. On the held-out test set, the proposed GAN achieved Dice = 0.873, IoU = 0.792, precision = 0.891, recall = 0.856, and Hausdorff distance = 4.2 pixels, outperforming a classical U-Net baseline (Dice = 0.814, IoU = 0.736), particularly in low-contrast and ambiguous regions. Additionally, the segmented fibrosis burden demonstrated strong clinical relevance through a negative correlation with left ventricular ejection fraction. These findings indicate that adversarially trained segmentation can improve objectivity and robustness of fibrosis assessment on routine TTE, supporting faster and more reproducible clinical decision-making.