EAGT: Echocardiography Augmentation for Generalisability and Transferability

2026-05-19 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Data Science & Analytics · Depth: Expert, extended

Summary

This study systematically evaluates 29 data augmentation techniques and their pairwise combinations to improve the cross-dataset generalizability of deep learning models for 2D left ventricular (LV) segmentation in echocardiography. Using a U-Net architecture, models were trained on Unity, CAMUS, and EchoNet Dynamic datasets and evaluated in-domain and across datasets using Dice and IoU scores. The research found that anatomically plausible geometric transformations, such as affine, shift-scale-rotate, perspective, and random horizontal flip, significantly enhance cross-dataset performance. Conversely, aggressive intensity- or artifact-based augmentations often degrade generalizability. Pairwise combinations, particularly moderate flip-centric ones like random horizontal flip with affine, consistently yielded greater gains than individual augmentations, providing empirically grounded guidance for designing robust augmentation policies for echocardiography segmentation models.

Key takeaway

For Computer Vision Engineers developing echocardiography segmentation models, you should prioritize geometric data augmentations like affine, shift-scale-rotate, perspective, and random horizontal flip to enhance cross-dataset generalizability. Combining random horizontal flip with affine transformations offers particularly robust performance improvements, especially when transferring models between different scanner vendors or acquisition protocols. Avoid overly aggressive intensity-based augmentations, as they can degrade performance in cross-dataset scenarios.

Key insights

Geometric data augmentations significantly improve echocardiography model generalizability across diverse datasets, outperforming intensity-based methods.

Principles

• Cross-dataset evaluation is crucial for assessing model generalizability.
• Anatomically plausible transformations are key for medical image augmentation.
• Pairwise augmentation combinations can yield synergistic performance gains.

Method

A U-Net model was trained on three echocardiography datasets, evaluating 29 individual and selected pairwise augmentation techniques with multiple hyperparameter settings, using Dice and IoU metrics and t-tests for statistical significance.

In practice

• Prioritize affine, shift-scale-rotate, perspective, and horizontal flip augmentations.
• Combine random horizontal flip with affine transformations for robust gains.
• Avoid aggressive intensity- or artifact-based augmentations for cross-dataset tasks.

Topics

• Data Augmentation
• Echocardiography Segmentation
• Model Generalisability
• Deep Learning
• Left Ventricular Segmentation

Code references

• soroushelyasi/augmentation_benchmark
• adamtupper/ultrasound-augmentation

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements — ROBUST-WT demonstrates that targeted training enhancements can substantially improve medical image segmentation robustness without altering core architecture.
• 1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization — 1S-DAug improves few-shot learning generalization by generating diverse, faithful image variants at test time via diffusion.
• Data Augmentation: A Fourier Analysis Perspective — Partial data augmentation can achieve full statistical benefits with significantly fewer samples than full group augmentation.
• Multi-Stage Bi-Atrial Segmentation Framework from 3D Late Gadolinium-Enhanced MRI using V-Net Family Models — A multi-stage V-Net framework with MCLAHE preprocessing effectively segments bi-atrial regions from 3D LGE-MRI.
• Exploring Data Augmentation and Resampling Strategies for Transformer-Based Models to Address Class Imbalance in AI Scoring of Scientific Explanations in NGSS Classroom — Targeted data augmentation significantly improves transformer-based model performance in automated scoring of imbalanced educational text data.
• Learning a Sampling-Free Variational DNN Plugin from Tiny Training Sets to Refine OOD Segmentation With Uncertainty Estimation — VarDeepPCA refines OOD medical image segmentations by learning anatomical geometric priors from small in-distribution datasets using sampling-free variational learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.