SkiNet: A deep learning framework for skin lesion diagnosis with uncertainty estimation and explainability

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer is the most common human malignancy, with millions of new cases annually.
• The limited dermatologist-to-patient ratio hinders early detection, especially in developing nations.
• Existing computer-aided diagnosis systems often lack the transparency required for medical practitioner trust.

Purpose of the Study:

• To introduce SkiNet, a novel deep learning architecture for rapid skin cancer screening.
• To develop an interpretable "white box" AI solution to improve trust in computer-aided diagnosis.
• To assist newly trained physicians in the clinical diagnosis of skin cancer.

Main Methods:

• A two-stage pipeline involving lesion segmentation and classification.
• Utilizing Monte Carlo dropout and test-time augmentation for uncertainty estimation.
• Employing a Bayesian MultiResUNet for segmentation uncertainty and saliency-based methods (XRAI, Grad-CAM, Guided Backprop) for model explanations.

Main Results:

• Demonstrated robustness of SkiNet on traditional benchmarks using the ISIC-2018 dataset.
• Successfully addressed the "black-box" nature of deep learning models in medical diagnosis.
• Incorporated transparency and confidence estimation into AI predictions for clinical use.

Conclusions:

• SkiNet provides a robust and interpretable solution for skin cancer screening.
• The model's transparency and uncertainty estimation can alleviate medical practitioner skepticism.
• This approach facilitates wider adoption of AI in clinical diagnosis, improving patient outcomes.