ExplainabilityTrustworthinessAccountabilitycs.CV
XAI for Skin Cancer Detection with Prototypes and Non-Expert Supervision
Miguel Correia, Alceu Bissoto, Carlos Santiago, Catarina Barata
February 2, 2024
arXiv.org
2 citations
1 influential
Key Takeaways
-
Interpretability can match or exceed black-box performance: The paper demonstrates that an interpretable AI model for melanoma detection achieves "superior performance and generalization compared to non-interpretable models," suggesting that transparency doesn't require sacrificing accuracy—a key concern for clinical adoption.
-
Non-expert input can guide AI systems effectively: By incorporating feedback from non-experts through binary masks and user-refined prototypes, the model learned to focus on relevant skin lesion features while excluding confounding factors, showing that expert-level supervision may not always be necessary for building trustworthy AI systems.
-
Prototype-based explanations make AI decisions visible to doctors: The prototypical-part model approach allows physicians to see which specific visual features the AI is using to make diagnoses, directly addressing the "black-box" concern that has made doctors hesitant to trust AI tools in clinical settings.
-
Automated guidance can reduce expert burden: The use of automatically-generated segmentation masks combined with user refinement creates a practical pathway for developing interpretable AI without requiring extensive expert annotation, potentially making responsible AI development more scalable in healthcare.
Abstract
Skin cancer detection through dermoscopy image analysis is a critical task. However, existing models used for this purpose often lack interpretability and reliability, raising the concern of physicians due to their black-box nature. In this paper, we propose a novel approach for the diagnosis of melanoma using an interpretable prototypical-part model. We introduce a guided supervision based on non-expert feedback through the incorporation of: 1) binary masks, obtained automatically using a segmentation network; and 2) user-refined prototypes. These two distinct information pathways aim to ensure that the learned prototypes correspond to relevant areas within the skin lesion, excluding confounding factors beyond its boundaries. Experimental results demonstrate that, even without expert supervision, our approach achieves superior performance and generalization compared to non-interpretable models.