Emerging Artificial Intelligence Models for Estimating Breslow Thickness from Dermoscopic Images.

Area of Science:

• Dermatology and Artificial Intelligence
• Medical Imaging Analysis

Background:

• Breslow thickness (BT) is crucial for melanoma prognosis but histopathological measurement has limitations.
• Preoperative clinical assessment shows significant misclassification rates.

Purpose of the Study:

• To review deep learning (DL) models for non-invasive Breslow thickness estimation from dermoscopic images.
• To assess the performance, limitations, and future directions of AI in melanoma BT assessment.

Main Methods:

• Systematic review of studies utilizing deep learning models (CNNs, Transformers) with transfer learning for BT estimation.
• Analysis of preprocessing techniques, interpretability methods, and external validation results.

Main Results:

• DL models achieve moderate accuracy (75-79%) and AUC (0.76-0.85) on single-center datasets.
• Performance degrades on external validation, with poor discrimination in critical thickness ranges and significant bias towards lighter skin tones.

Conclusions:

• AI models can serve as decision-support tools but do not replace histopathology.
• Further research is needed for diverse datasets, improved accuracy at critical thresholds, multi-institutional validation, and regulatory pathways.