Predicting the clinical management of skin lesions using deep learning

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Automated machine learning (ML) for skin lesion diagnosis nears dermatologist performance.
• Current ML management suggestions infer decisions from predicted diagnoses, ignoring decision variability.
• This overlooks the nuanced decision-making process in clinical practice.

Purpose of the Study:

• To develop and evaluate an ML model for direct image-based prediction of skin lesion management decisions.
• To compare the accuracy of direct management prediction against diagnosis-inferred management.
• To assess the impact on reducing unnecessary excisions and model generalizability.

Main Methods:

• Utilized clinical and dermoscopic images with patient metadata from the Interactive Atlas of Dermoscopy dataset (1011 cases).
• Developed ML models to predict management decisions directly from images, bypassing explicit diagnosis prediction.
• Incorporated simultaneous prediction of seven-point criteria and diagnosis for enhanced management prediction accuracy.
• Evaluated model generalizability on the MClass-D dataset.

Main Results:

• Direct management prediction significantly outperformed diagnosis-inferred management (improved accuracy and AUROC).
• The direct approach reduced the over-excision rate by 24.56%.
• Simultaneous prediction of diagnosis and seven-point criteria further improved management prediction accuracy.
• The model demonstrated strong generalizability, aligning with dermatologist consensus on external datasets.

Conclusions:

• Directly predicting clinical management decisions for skin lesions from images is more accurate and efficient than inferring from diagnoses.
• This approach has the potential to significantly reduce unnecessary surgical procedures.
• Integrating diagnostic criteria alongside management prediction enhances model performance and clinical utility.