Identifying melanoma among benign simulators - Is there a role for deep learning convolutional neural networks?...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Early detection of cutaneous melanoma (CM) is vital for survival, but distinguishing it from benign melanoma simulators (MelSim) is challenging.
• Deep learning convolutional neural networks (DL-CNNs) show promise in identifying CM with dermatologist-level accuracy.

Purpose of the Study:

• To evaluate if DL-CNN support enhances dermatologists' accuracy in differentiating CM from MelSim.
• To compare diagnostic performance with and without DL-CNN assistance.

Main Methods:

• A cross-sectional reader study involving 200 skin lesions (100 CM, 100 MelSim) and 27 dermatologists.
• Dermatologists assessed lesions using dermoscopy alone, full case information, and full case information plus DL-CNN scores.
• Diagnostic performance metrics included sensitivity, specificity, and ROC-AUC.

Main Results:

• DL-CNN and dermatologists (without support) had comparable sensitivity (90.0% vs 90.1%).
• DL-CNN had lower specificity (67.0%) and ROC-AUC (0.889) than dermatologists (73.2% and 0.951, respectively).
• With DL-CNN support, dermatologists' sensitivity increased to 91.4% (p < 0.001) without significant changes in specificity or ROC-AUC.

Conclusions:

• Collaboration with DL-CNNs slightly improved dermatologists' diagnostic accuracy for CM.
• The enhancement was primarily due to increased sensitivity, with no significant loss in specificity.
• DL-CNNs show potential as valuable assistant diagnostic tools in challenging cases.