DEA: Data-efficient augmentation for interpretable medical image segmentation

Data efficiency plays a pivotal role in medical image segmentation where data labeling is expensive and time consuming. However, there are few effective methods to enhance data efficiency without compromising the methods' effectiveness. We propose a plug-and-use data augmentation method called Data Efficient Augmentation (DEA) for efficient medical image segmentation. DEA is designed to enhance data efficiency, which also has excellent generalization capabilities across different segmentation methods without modifying network structures. Furthermore, we extract data-efficient features with Grad-weighted Class Activation Mapping (Grad-CAM). The proposed DEA method improves segmentation performance on Hyper-Kvasir and ISIC-Archive datasets. With the proposed DEA method, the Intersection over Union (IoU) and dice similarity coefficient (Dice) is increased by 2.84% and 2.26% respectively compared with the state-of-the-art methods. What is more, DEA enables different segmentation methods to achieve over 95% accuracy with only 70% of training data compared with these methods using the whole dataset. The quantitative and qualitative analysis proves that the proposed DEA is tailored for medical image segmentation and improves the interpretability in data augmentation techniques.

Automated summary

Data Efficient Augmentation (DEA) is proposed as a plug-and-use method for efficient medical image segmentation. DEA enhances data efficiency and has good generalization capabilities across different segmentation methods, improving segmentation performance on multiple datasets.