Soroush Elyasi, Sara Adibzadeh, Nasim Dadashi Serej et al.
Deep learning models for echocardiography segmentation often struggle to generalise across institutions, scanners, and patient populations, where collecting large, consistently annotated datasets is infeasible. Data augmentation is widely used to improve the robustness of deep learning models; however, its role in enhancing cross-dataset generalisability in echocardiography remains insufficiently understood. This study presents a large-scale multi-dataset evaluation of 29 data augmentation techniques and their pairwise combinations for 2D left ventricular segmentation using a U-Net trained on Unity, CAMUS, and EchoNet Dynamic datasets. Each augmentation was explored under several hyperparameter settings and assessed through repeated runs using Dice and IoU in both in-domain and cross-dataset scenarios, with statistical significance quantified via independent t-tests. Results show that anatomically plausible geometric transformations, particularly affine, shift-scale-rotate, perspective, and random horizontal flip, substantially improve cross-dataset performance, whereas aggressive intensity- or artefact-based augmentations often degrade generalisability. Pairwise augmentation combinations outperform individual augmentations and show that moderate flip-centric combinations, especially random horizontal flip with affine, yield consistent gains across most transfer scenarios. These findings provide empirically grounded guidance for designing augmentation policies that enhance the robustness and transferability of echocardiography segmentation models.