Towards the Generation of Medical Imaging Classifiers Robust to Common Perturbations

Abstract

Background: Machine learning (ML) and artificial intelligence (AI)-based classifiers can be used to diagnose diseases from medical imaging data. However, few of the classifiers proposed in the literature translate to clinical use because of robustness concerns. Materials and methods: This study investigates how to improve the robustness of AI/ML imaging classifiers by simultaneously applying perturbations of common effects (Gaussian noise, contrast, blur, rotation, and tilt) to different amounts of training and test images. Furthermore, a comparison with classifiers trained with adversarial noise is also presented. This procedure is illustrated using two publicly available datasets, the PneumoniaMNIST dataset and the Breast Ultrasound Images dataset (BUSI dataset). Results: Classifiers trained with small amounts of perturbed training images showed similar performance on unperturbed test images compared to the classifier trained with no perturbations. Additionally, classifiers trained with perturbed data performed significantly better on test data both perturbed by a single perturbation (p-values: noise = 0.0186; contrast = 0.0420; rotation, tilt, and blur = 0.000977) and multiple perturbations (p-values: PneumoniaMNIST = 0.000977; BUSI = 0.00684) than the classifier trained with unperturbed data. Conclusions: Classifiers trained with perturbed data were found to be more robust to perturbed test data than the unperturbed classifier without exhibiting a performance decrease on unperturbed test images, indicating benefits to training with data that include some perturbed images and no significant downsides.