MedAugment: Universal Automatic Data Augmentation Plug-in for Medical Image Analysis

TL;DR

MedAugment addresses data scarcity in medical image analysis by introducing a dual-space automatic data augmentation framework with constrained sampling and a single-parameter hyperparameter mapping. The two spaces, $A_p$ (pixel) and $A_s$ (spatial), together with a mapping from augmentation level $l\in\{1,2,3,4,5\}$, enable realistic, diagnostically faithful augmentations while preserving computational efficiency and offering plug-in compatibility. Across eight public datasets for classification and segmentation and multiple architectures, MedAugment consistently outperforms state-of-the-art DA methods such as StyleGAN2-ADA, AutoAugment, RandAugment, and AugMix, demonstrating robust gains in accuracy and segmentation quality. The approach is designed to be accessible to clinicians and adaptable to arbitrary projects without extra training, highlighting its practical impact on medical image analysis, though future work could further tailor augmentations to specific datasets and address small-object segmentation challenges.

Abstract

Data augmentation (DA) has been widely leveraged in computer vision to alleviate the data shortage, whereas the DA in medical image analysis (MIA) faces multiple challenges. The prevalent DA approaches in MIA encompass conventional DA, synthetic DA, and automatic DA. However, utilizing these approaches poses various challenges such as experience-driven design and intensive computation cost. Here, we propose an efficient and effective automatic DA method termed MedAugment. We propose a pixel augmentation space and spatial augmentation space and exclude the operations that can break medical details and features, such as severe color distortions or structural alterations that can compromise image diagnostic value. Besides, we propose a novel sampling strategy by sampling a limited number of operations from the two spaces. Moreover, we present a hyperparameter mapping relationship to produce a rational augmentation level and make the MedAugment fully controllable using a single hyperparameter. These configurations settle the differences between natural and medical images, such as high sensitivity to certain attributes such as brightness and posterize. Extensive experimental results on four classification and four segmentation datasets demonstrate the superiority of MedAugment. Compared with existing approaches, the proposed MedAugment serves as a more suitable yet general processing pipeline for medical images without producing color distortions or structural alterations and involving negligible computational overhead. We emphasize that our method can serve as a plugin for arbitrary projects without any extra training stage, thereby holding the potential to make a valuable contribution to the medical field, particularly for medical experts without a solid foundation in deep learning. Code is available at https://github.com/NUS-Tim/MedAugment.

Figures (8)

• Figure 1: A realization of MedAugment. The MedAugment consists of $N = 4$ augment branches and a separate branch to retain the input features. For each branch, $M = {\{2, 3}\}$ DA operations are sampled using the sampling strategy $\Pi$ from the pixel augmentation space $A_{p}$ and spatial augmentation space $A_{s}$.
• Figure 2: Examples of augmented medical images generated by varying automatic DA methods.
• Figure 3: Class activation map across different DA methods on the BUSI dataset using VGGNet. The red regions present the region of interest.
• Figure 4: T-SNE visualization across different DA methods on the BUSI dataset using VGGNet.
• Figure 5: Predicted and GT masks across different DA methods on varying datasets using DeepLabV3.