Data Augmentation in Training Deep Learning Models for Medical Image Analysis

Abstract

Data augmentation is widely utilized to achieve more generalizable and accurate deep learning models based on relatively small labeled datasets. Data augmentation techniques are particularly critical in medical applications, where access to labeled data samples is commonly limited. Although data augmentation methods generally have a positive impact on the performance of deep learning models, not all data augmentation techniques are applicable and suitable for analyzing medical images. In this chapter, we review common image augmentation techniques and their properties. Furthermore, we present and evaluate application-specific data augmentation methods that are beneficial for medical image analysis. The material presented in this chapter aims to guide the use of data augmentation techniques in training deep learning models for various medical image analysis applications, in which annotated data are not abundant or are difficult to acquire.