Machine Learning in Melanoma Diagnosis. Limitations About to be Overcome

Area of Science:

• Dermatology and Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Automated image classification using machine learning (ML) shows promise for skin cancer diagnosis.
• Current limitations hinder the general usability of ML in clinical practice for skin cancer detection.

Purpose of the Study:

• To identify and quantify limitations in selecting skin cancer images for ML analysis, with a focus on melanoma.
• To assess the eligibility of dermoscopy images for ML-based diagnostic systems.

Main Methods:

• A retrospective cohort study analyzed 2,849 dermoscopy images of skin tumors (2010-2014).
• Images were evaluated for eligibility criteria for ML analysis.
• Exclusion criteria and reasons for exclusion were systematically recorded.

Main Results:

• Only 34% (968/2849) of all images met the inclusion criteria for ML analysis.
• Melanoma image eligibility was low, with only 36.6% meeting criteria.
• Common exclusion reasons for melanoma images included absence of normal surrounding skin (40.5%) and lack of pigmentation (14.2%).

Conclusions:

• A significant majority of melanoma images (63.4%) were excluded from ML analysis.
• Current state-of-the-art ML systems require training on larger, more diverse datasets that include non-ideal images from real clinical settings.
• Ongoing research is actively addressing these limitations for improved clinical applicability.