Super-Trustscore: Reliable Failure Detection for Automated Skin Lesion Diagnosis
Abstract: The successful deployment of deep neural networks in safety-critical settings, such as medical image analysis, is contingent on their ability to provide reliable uncertainty estimates. In this paper, we propose a new confidence scoring function called Super-TrustScore that improves upon the existing TrustScore method by combining a local confidence score and a global confidence score. Super-TrustScore is a post-hoc method and can be easily applied to any existing pre-trained model as there are no particular architecture or classifier training requirements. We demonstrate empirically that Super-TrustScore consistently provides the most reliable uncertainty estimates for both in-distribution and shifted-distribution failure detection on the task of skin lesion diagnosis.
Loading