Fair Conformal Predictors for Applications in Medical Imaging

Deep learning has the potential to augment several clinically useful aspects of the radiologist's workflow such as medical imaging interpretation. However, the translation of deep learning algorithms into clinical practice has been hindered by relative lack of transparency in these algorithms compared to more traditional statistical methods. Specifically, common deep learning models lack intuitive and rigorous methods of conveying prediction confidence in a calibrated manner, which ultimately restricts widespread use of these "black box" systems for critical decision-making. Furthermore, numerous demonstrations of algorithmic bias in clinical machine learning have caused considerable hesitancy towards the deployment of these models for clinical application. To this end, we explore how conformal predictions can complement existing deep learning approaches by providing an intuitive way of expressing model uncertainty to facilitate greater transparency to clinical users. In this paper, we conduct field interviews with radiologists to assess potential use-cases of conformal predictors. Using insights collected from these interviews, we devise two use-cases and empirically evaluate several conformal methods on a dermatology photography dataset for skin lesion classification. Additionally, we show how group conformal predictors are more adaptive to differences between patient skin tones for malignant skin lesions. We find our conformal predictors to be a promising and generally applicable approach to increasing clinical usability and trustworthiness -- hopefully facilitating better modes of collaboration between medical AI tools and their clinical users.