SANO: Score-Based Diffusion Model for Anomaly Localization in Dermatology
Supervised learning for dermatology requires a large volume of annotated images, but collecting clinical data is costly, and it is virtually impossible to cover all clinical cases. Unsupervised anomaly localization circumvents this problem by learning the healthy data distribution. However, algorithms which use a generative model and localize pathologic regions based on a reconstruction error are not robust to domain shift, which is a problem for dermatology due to the low level of standardization expected in many applications. Our method, SANO, uses score-based diffusion models to produce a log-likelihood gradient map highlighting areas that contain abnormalities. A segmentation mask can then be calculated based on deviations from typical values observed during training. After benchmarking SANO on an industrial dataset, we train it on a public non-clinical dataset of healthy hand images without ornaments, evaluate it on the task of detecting jewelry within images from the same dataset, and prove its robustness by using it on clinical pictures to localize hand eczema. We demonstrate that SANO outperforms competing approaches from the literature without introducing additional computational costs.