Cardiac Ultrasonic Tissue Characterization in Myocardial Infarction Based on Deep Transfer Learning and Radiomics Features

Abstract

Objective: Acute myocardial infarction (MI) alters cardiomyocyte geometry and architecture, leading to changes in the acoustic properties of the myocardium. This study examines ultrasomics - a novel cardiac ultrasound-based radiomics technique to extract high-throughput pixel-level information from images - for identifying infarcted myocardium. Methodology: A retrospective multicenter cohort of 380 participants was split into two groups: a model development cohort (n=296; 101 MI cases, 195 controls) and an external validation cohort (n=84; 40 MI cases, 44 controls). Handcrafted and transfer learning-derived deep ultrasomics features were extracted from 2-chamber and 4-chamber echocardiographic views and ML models were built to detect patients with MI and infarcted myocardium within individual views. Myocardial infarct localization via texture features was determined using Shapley additive explanations. All the ML models were trained using 10-fold cross-validation and assessed on an external test dataset, using the area under the curve (AUC). Results: The ML model, leveraging segment-level handcrafted ultrasomics features identified MI with AUCs of 0.93 (95% CI: 0.97-0.97) and 0.83 (95% CI: 0.74-0.89) at the patient-level and view-level, respectively. A model combining handcrafted and deep ultrasomics provided incremental information over deep ultrasomics alone (AUC: 0.79, 95% CI: 0.71-0.85 vs. 0.75, 95% CI: 0.66-0.82). Using a view-level ultrasomic model we identified texture features that effectively discriminated between infarcted and non-infarcted segments (p<0.001) and facilitated parametric visualization of infarcted myocardium. Conclusion: This pilot study highlights the potential of cardiac ultrasomics in distinguishing healthy and infarcted myocardium and opens new opportunities for advancing myocardial tissue characterization using echocardiography.