Melanoma recognition by a deep learning convolutional neural network-Performance in different melanoma subtypes and...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning convolutional neural networks (CNNs) demonstrate significant promise for melanoma diagnosis.
• Melanoma thickness, influenced by location and subtype, presents diagnostic challenges.
• The ability of CNNs to address these challenges, particularly in difficult subtypes, remains under-investigated.

Purpose of the Study:

• To evaluate the diagnostic performance of a European market-approved CNN across various melanoma localizations and subtypes.
• To determine if CNNs can assist physicians in diagnosing challenging melanoma cases.

Main Methods:

• A commercial CNN (Moleanalyzer-Pro®) was utilized for binary classification (malignant/benign) on six distinct dermoscopic image sets.
• Each set comprised 30 melanomas and 100 benign lesions, representing specific localizations and morphologies (superficial spreading, lentigo maligna, nodular, mucosal, acrolentiginous skin, and subungual).

Main Results:

• The CNN achieved high performance (sensitivity >93.3%, specificity >65%, ROC-AUC >0.926) in superficial spreading, nodular, and lentigo maligna melanomas.
• For acrolentiginous melanomas of the skin, performance was good (sensitivity 83.3%, specificity 91.0%, ROC-AUC 0.928).
• Limited diagnostic performance was observed for mucosal melanomas (specificity 38.0%) and subungual melanomas (sensitivity 53.3%, specificity 68.0%, ROC-AUC 0.621).

Conclusions:

• CNNs can partially compensate for reduced human diagnostic accuracy in melanoma detection.
• Physicians must be aware of the CNN's limitations in diagnosing mucosal and subungual melanomas.
• Enhanced CNN performance in these challenging sites may be achieved through additional, targeted training data.