[Paper Review] Analyzing Data Augmentation for Medical Images: A Case Study in Ultrasound Images

Adam Tupper, Christian Gagné|arXiv (Cornell University)|Mar 14, 2024

Medical Imaging Techniques and ApplicationsMedicine3 citations

TL;DR

This study rigorously evaluates data augmentation for breast lesion classification in ultrasound images, demonstrating that randomly sampling from a diverse pool of augmentations—particularly using TrivialAugment—yields up to 10.4% accuracy gains over baseline methods. The approach consistently outperforms fixed sequences and individual augmentations, with geometric transforms proving more effective than photometric ones.

ABSTRACT

Data augmentation is one of the most effective techniques to improve the generalization performance of deep neural networks. Yet, despite often facing limited data availability in medical image analysis, it is frequently underutilized. This appears to be due to a gap in our collective understanding of the efficacy of different augmentation techniques across medical imaging tasks and modalities. One domain where this is especially true is breast ultrasound images. This work addresses this issue by analyzing the effectiveness of different augmentation techniques for the classification of breast lesions in ultrasound images. We assess the generalizability of our findings across several datasets, demonstrate that certain augmentations are far more effective than others, and show that their usage leads to significant performance gains.

Motivation & Objective

• To address the underutilization of data augmentation in medical imaging, particularly in breast ultrasound classification.
• To investigate why certain augmentation techniques are more effective than others across different datasets and tasks.
• To determine whether fixed sequences of augmentations or random sampling from a diverse set yields better performance.
• To provide a standardized, empirically validated data augmentation strategy for deep learning in breast ultrasound imaging.

Proposed method

• The study employs a $5\times2$ cross-validation framework with paired t-tests and Holm-Bonferroni correction to ensure statistical rigor in comparing augmentation strategies.
• Individual augmentation effectiveness is evaluated using one-sided t-tests on commonly used configurations, focusing on significant performance improvements.
• Paired augmentation performance is assessed by testing ordered combinations of two transforms, examining interaction effects and compound gains.
• Random sampling strategies are evaluated by increasing the size of the augmentation pool and the number of operations applied per image using TrivialAugment.
• The analysis compares performance across three distinct tasks: BUSI pathology, BUS-BRA pathology, and BUS-BRA BI-RADS classification.
• Photometric and geometric augmentations are evaluated separately and in combination to assess their relative contributions to model performance.

Figure 1 : The change in balanced validation accuracy for each data augmentation across the three tasks.

Experimental results

Research questions

• RQ1 Which individual data augmentation techniques significantly improve classification performance for breast lesions in ultrasound images across different datasets?
• RQ2 How does the performance of fixed sequences of augmentations compare to random sampling from a diverse set of augmentations?
• RQ3 Does the effectiveness of individual augmentations vary across different tasks (e.g., pathology vs. BI-RADS classification) and datasets?
• RQ4 What is the impact of combining photometric and geometric augmentations on model generalization and accuracy?
• RQ5 Can a random sampling strategy from a diverse augmentation pool consistently outperform deterministic sequences, even when individual augmentations are weak?

Key findings

• Random sampling from a diverse pool of augmentations using TrivialAugment led to a 10.4% accuracy gain on the BUS-BRA BI-RADS classification task, the highest improvement observed.
• The only individual augmentation that significantly improved performance across multiple tasks was random rotation, highlighting task- and dataset-specific effectiveness.
• Fixed sequences of augmentations provided minimal benefit and often failed to yield substantial gains, even when combining strong transforms.
• Using geometric transforms alone outperformed photometric transforms, but combining both yielded the best overall performance.
• Performance gains increased with both the size of the augmentation pool and the number of operations applied, indicating that diversity and multiplicity are more important than individual strength.
• The strategy of random sampling from a diverse set of augmentations consistently outperformed all other methods, even when individual augmentations were ineffective.

Figure 2 : Change in balanced validation accuracy using two sequential operations compared to using Operation 1 alone.