Decoding Diagnosis: AI Explainability for Enhanced Skin Cancer Detection.

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer diagnosis relies heavily on visual examination, with AI showing promise in classification.
• Deep learning models for AI diagnosis lack transparency, hindering dermatologist trust and interpretability.
• Prototypical Part Networks (ProtoPNet) offer a framework for interpretable AI in medical diagnosis.

Purpose of the Study:

• To investigate the effectiveness of Pixel-Grounded Prototypical Part Networks (PIXPNET) for skin cancer detection.
• To compare the performance and interpretability of PIXPNET against the original ProtoPNet.
• To address the challenge of pixel-space mapping in prototype projection for enhanced AI interpretability.

Main Methods:

• Applied the PIXPNET architecture, an advancement of ProtoPNet, to skin cancer detection tasks.
• Trained and evaluated the PIXPNET model on a multi-class classification dataset.
• Analyzed learned prototypes within PIXPNET to assess their relevance and interpretability in relation to input images.

Main Results:

• PIXPNET demonstrated significantly superior performance compared to ProtoPNet in multi-class skin cancer classification.
• The model's generalizability was successfully assessed through training and evaluation.
• Learned prototypes in PIXPNET showed improved relevance and interpretability when compared to ProtoPNet.

Conclusions:

• PIXPNET represents a significant advancement in interpretable AI for skin cancer detection.
• The proposed model enhances diagnostic transparency, aiding dermatologists in understanding AI-driven classifications.
• PIXPNET offers a more interpretable and accurate alternative to existing AI methods in dermatological applications.