Fortifying medical image segmentation: adversarially robust and trustworthy deep learning solutions

Abstract

In the era of digital innovation, Deep Learning (DL) has achieved remarkable success across various fields, including healthcare, where its decisions directly impact human lives. The DL-based medical models are versatile and capable of handling a wide range of tasks. Medical Image Segmentation (MIS) is one such critical task wherein a disease-afflicted Region of Interest (ROI) is separated from the unintended regions. Such ROIs primarily refer to any organs, cancerous cells, tissues, or lesions. Unfortunately, the DL-based MIS models are highly vulnerable to intelligently curated adversarial attacks, where small and imperceptible perturbations are imposed on the input that drastically mislead the predictions. This concern is more pervasive with the medical images, as their rich textural details can easily divert the focus of the model towards irrelevant regions, ultimately diminishing their performance and robustness. Moreover, the predictions of such models struggle with anomalous samples like adversarial and Out-Of-Distribution (OOD). While adversarial samples are generated by adding small and imperceptible perturbations to the input, OOD samples signify input data with a shifted distribution. Both of these samples significantly degrade the performance of the model. Furthermore, the opaque nature of DL models offers non-trustworthy predictions, causing conflicts among users over accepting the model prediction. All these problems could result in catastrophic consequences in the healthcare domain.