Artificial intelligence for melanoma diagnosis.

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) demonstrate high accuracy in digital image analysis, including clinical and dermatoscopic images for melanoma recognition.
• Modern CNN architectures achieve expert-level performance in single image analysis and multiclass diagnoses in experimental settings.
• Despite experimental success, reliable prospective clinical trials for CNNs in dermatology are lacking, and their translation to clinical practice is uncertain.

Purpose of the Study:

• To evaluate the potential and limitations of Convolutional Neural Networks (CNNs) in automated melanoma recognition and histologic assessment.
• To explore the current state of CNN application in dermatology, focusing on diagnostic accuracy and clinical utility.
• To identify challenges and future directions for CNN integration into dermatologic practice, particularly for complex lesions.

Main Methods:

• Review of current Convolutional Neural Network (CNN) architectures and their performance in analyzing clinical and dermatoscopic images for skin lesion classification.
• Analysis of experimental study findings comparing CNN performance to dermatologists and domain experts.
• Discussion of limitations including training dataset biases, uncertainty estimation, and the need for prospective clinical validation.

Main Results:

• CNNs exhibit high accuracy in experimental image analysis, comparable to human experts for melanoma recognition and differential diagnoses.
• Limitations of training datasets directly impact CNN prediction accuracy and generalizability.
• CNNs currently lack reliable uncertainty estimation capabilities, a critical factor for clinical decision-making.
• Human-computer collaboration with CNNs has shown potential for improved diagnostic accuracy, even with imperfect models.

Conclusions:

• While CNNs show promise for melanoma recognition, their direct clinical application is hindered by data limitations, lack of uncertainty quantification, and insufficient clinical validation.
• Fully automating histologic assessment of equivocal melanocytic lesions using CNNs is currently problematic due to diagnostic ambiguities and lack of consensus.
• Near-future applications of CNNs in dermatology are likely to be supportive, serving roles in referencing, recommendations, or aiding clinical decision-making rather than full automation.