A System for the Analysis of Dermoscopy Images Using Weak Annotations

This paper proposes a two-step approach for the analysis of dermoscopy images. In the first step, we detected dermoscopic criteria (structures and colors), which are used by dermatologists in their medical analysis. In the second step, this information is used to automatically diagnose skin cancer. The extraction of dermoscopic criteria from skin lesions is a challenging task because the amount of detailed annotated images is scarce. We solve this task by using a probabilistic model (topic model) learned from weakly annotated data. This approach overcomes the need for completely annotated datasets, only requiring text labels. The second step uses the detected criteria to train a Random Forest classifier. The system achieves a good classification score: sensitivity of 85.8% and a specificity of 71.1%. Nonetheless, the main advantage of this system with respect to others is its ability to justify the decision based on medical criteria.