Robust Training with Data Augmentation for Medical Imaging Classification

Deep learning (DL) has shown significant promise in medical imaging applications, especially in the detection and diagnosis of conditions such as breast cancer through mammogram analysis [1]. Despite their increasing adoption, deep neural networks (DNNs) remain vulnerable to adversarial attacks and natural variations. A meta-analysis of 516 published research works showed that less than 6% of AI applications for the diagnostic analysis of medical images incorporated external validation, such as distribution shift tests [15]. Furthermore, according to the FUTURE AI guidelines [16], improving the generalizability and robustness of computer-aided diagnostic systems is essential, particularly in healthcare applications where model safety and reliability are paramount. These vulnerabilities are particularly concerning in critical applications such as healthcare, where errors can have profound consequences for patient outcomes. Adversarial attacks refer to carefully crafted perturbations designed to manipulate the predictions of the model.