Super-trustscore: reliable failure detection for automated skin lesion diagnosis
The successful deployment of deep neural networks in safetycritical 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...
| Huvudupphovsmän: | , |
|---|---|
| Materialtyp: | Conference item |
| Språk: | English |
| Publicerad: |
IEEE
2024
|
| Sammanfattning: | The successful deployment of deep neural networks in safetycritical 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 SuperTrustScore consistently provides the most reliable uncertainty
estimates for both in-distribution and shifted-distribution failure detection on the task of skin lesion diagnosis. |
|---|