One model to segment them all: generalisation vs specificity in echocardiography left ventricle delineation

Introduction Accurate segmentation of left ventricle (LV) is essential for quantitative echocardiographic analysis and clinical decision-making. Deep learning models have proven effective in automating this task, often relying on view-specific data to improve accuracy [1]. Nevertheless, this reliance limits scalability, increases complexity of model development, and overlooks shared anatomical features across views, ultimately reducing generalisability. Purpose Establishing LV echocardiography segmentation in apical views as case scenario, this study aims to compare the conventional approach of training multiple view-specific models to training one multi-view model. Methods A total of 153 studies from a cohort of 122 patients (mean age 64.8 ± 12.6 years; 40% female) were retrospectively considered for this work. LV endocardial and epicardial borders were manually delineated at end-systole and end-diastole in apical 2-chamber (A2C), 3-chamber (A3C), and 4-chamber (A4C) views by a cohort of 21 expert sonographers (mean experience 5 ± 3 years). A total of 707 frames were annotated (A2C: 250; A3C: 200; and A4C: 257). Two model development approaches were considered: training a single segmentation model for all views (multi-view model); and training an individual model for each specific view (single-view models). The nnU-Net framework [2] was employed in both strategies and the models were trained on a NVIDIA A100 40GB. Segmentation performance was assessed via Dice similarity coefficient (DSC), evaluating overlap between model predictions and manual ground truth. Results The multi-view model achieved an overall performance comparable to the single-view models (average DSC: 0.898±0.015 vs 0.896±0.015, respectively) (Figure 1). Although no statistical differences were observed at view or structure level, the multi-view model showed a slightly higher median and average DSC in the A2C and A3C views (Figure 2). Conversely, the A4C-specific model demonstrated a minor improvement in endocardium delineation. Both approaches reported comparable performance to state-of-the-art echocardiography segmentation methods [3]. Conclusion The proposed multi-view model approach demonstrates comparable LV segmentation performance to single-view models. This multi-view approach has the potential to ease model use and deployment by speeding-up development (1 vs. 3 models), simplifying workflow and eliminating dependencies on prior steps (e.g. view identification), thereby lowering computational demands and costs. Future work will focus on further assessing the generalisability as well as the model ability to capture the inherent cardiac structure.