Study Evaluates 29 Data Augmentation Techniques for Echocardiography Segmentation
A comprehensive study investigates 29 techniques for data augmentation and their various combinations for 2D left ventricular segmentation, utilizing a U-Net model trained on the Unity, CAMUS, and EchoNet Dynamic datasets. The goal of this research is to enhance cross-dataset generalizability in echocardiography, where obtaining extensive annotated datasets is challenging. Each augmentation technique underwent testing with different hyperparameter configurations and was evaluated using Dice and IoU metrics in both in-domain and cross-dataset contexts, with independent t-tests applied to determine statistical significance. Findings reveal that geometrically plausible anatomical transformations yield the most effective results.
Key facts
- 29 data augmentation techniques and their pairwise combinations were evaluated
- U-Net model trained on Unity, CAMUS, and EchoNet Dynamic datasets
- 2D left ventricular segmentation task
- Assessed using Dice and IoU metrics
- In-domain and cross-dataset scenarios tested
- Statistical significance via independent t-tests
- Anatomically plausible geometric transformations were most effective
- Study addresses generalisability across institutions, scanners, and patient populations
Entities
—