Revisiting Data Augmentation for Ultrasound Images

Adam Tupper, Christian Gagné

Abstract

Data augmentation (DA) is a cornerstone of modern deep learning, yet its application in medical ultrasound remains inconsistent and under-researched. This paper introduces UltraBench, a comprehensive benchmark comprising 14 diverse ultrasound tasks, to systematically evaluate the impact of various DA strategies. Our findings reveal that standard "natural image" augmentations often outperform specialized ultrasound-specific techniques. Furthermore, we demonstrate that automated strategies like TrivialAugment can significantly enhance performance, though their efficacy is highly task-dependent and subject to a "dilution effect."

Introduction

Despite the chronic shortage of labeled medical data, data augmentation is surprisingly underutilized in ultrasound research. Practitioners often hesitate to apply aggressive transformations, fearing that "natural image" augmentations—such as those used in ImageNet training—might create anatomically impossible artifacts that degrade model performance. Consequently, many studies either omit DA entirely or rely on a limited set of "realistic" transformations.

This study addresses the lack of a rigorous, large-scale comparison of DA techniques across diverse ultrasound tasks. By introducing UltraBench, we provide a standardized framework to evaluate whether specialized ultrasound-specific noise models truly offer advantages over simple geometric transforms.

The UltraBench Framework

To ensure a robust evaluation, we curated UltraBench, a collection of 14 tasks (7 classification and 7 segmentation) derived from 10 public data sources. The benchmark spans 11 body regions and includes:

  • Liver Analysis: Liver mass classification and segmentation (AUL & Fatty Liver).
  • Cardiac Imaging: Structure segmentation and quality classification (CAMUS).
  • Gynecology: Ovarian tumor analysis (MMOTU).
  • Multi-organ & Point-of-Care: Carotid, bladder, and lung ultrasound (Butterfly, POCUS).
  • Specialized Anatomy: Gallbladder (GBCU), Kidney (Open Kidney), Thyroid (Stanford Thyroid), and Fetal Head (PSFHS).

Methodology

Our systematic study followed a three-tiered approach:

  1. Individual Evaluation: We tested 14 augmentations in isolation. This included 10 "traditional" transforms (e.g., rotation, horizontal flipping, brightness adjustment, zoom, and random cropping) and 4 "ultrasound-specific" transforms (depth attenuation, haze artifact, Gaussian shadow, and speckle reduction).
  2. Automated Strategy Evaluation: We employed TrivialAugment, a strategy that randomly selects augmentations from a predefined pool. We specifically investigated how performance fluctuates as the pool expands from high-performing to lower-performing individual augmentations.
  3. Standardized Architectures: To isolate the effects of DA, we utilized EfficientNet-B0 for classification and UNet (with an EfficientNet backbone) for segmentation. Findings were subsequently validated using larger models, including EfficientNet-B5 and SegFormer.

Key Insights and Results

The Superiority of "Natural" Transforms

One of the most striking findings is that traditional geometric transforms—specifically Zoom and Random Crop—were the only augmentations to improve performance across every classification task. These "natural" transforms frequently outperformed niche augmentations engineered to mimic ultrasound physics, such as speckle reduction.

Task-Modality Sensitivity

The effectiveness of DA is highly sensitive to the specific task and organ. For instance, while Gaussian noise improved liver segmentation by 2.77%, it caused a catastrophic 40.41% decline in liver mass segmentation. This divergence underscores that a "one-size-fits-all" augmentation strategy does not exist, even when imaging the same organ.

Classification vs. Segmentation

We observed a significant disparity in how different task types respond to DA. Classification tasks saw substantial gains, with improvements ranging from 0.99% to 30.26% in mean Average Precision (mAP) over the baseline. In contrast, segmentation tasks showed more modest improvements, typically yielding increases between 0.38% and 9.02% in Dice scores. This suggests a lower performance ceiling for DA in pixel-level tasks compared to image-level classification.

The Dilution Effect

In testing TrivialAugment, we discovered that more is not always better. Performance typically peaked when the augmentation pool contained the top 5-9 transforms. Adding further, less effective transforms led to a "dilution effect," where the frequency of high-impact augmentations decreased, ultimately hindering the model's learning process.

Conclusion

This research challenges the prevailing caution regarding "natural" data augmentation in ultrasound. While ultrasound-specific transforms like depth attenuation can be highly effective in certain niches (e.g., an 8.74% boost in Fatty Liver classification), standard geometric transforms remain the most reliable baseline.

Our study was primarily limited to 2D "fan-shape" images from convex and phased array probes. Future research should extend these benchmarks to 3D ultrasound and linear probe geometries. By releasing UltraBench and our associated code, we aim to provide the community with the tools necessary to move beyond heuristic-based DA and toward evidence-based model training.

Summary updated

BibTeX
@article{tupper2025revisiting,
  author = {Adam Tupper and Christian Gagné},
  title = {{Revisiting Data Augmentation for Ultrasound Images}},
  journal = {Accepted by TMLR},
  year = {2025},
  url = {https://openreview.net/forum?id=iGcxlTLIL5},
}
Figure