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

## TL;DR

This paper reviews AI models for estimating Breslow thickness from dermoscopic images, highlighting their potential and limitations in clinical practice.

## Contribution

The paper introduces research priorities for improving AI fairness and accuracy in melanoma diagnosis across diverse skin types.

## Key findings

- Deep learning models achieve 75–79% accuracy in BT estimation but show poor discrimination in the 0.4–1.0 mm range.
- Most AI models are trained on lighter skin types, leading to potential bias against darker skin phototypes.
- AI is proposed as a complementary tool to histopathology, not a replacement.

## Abstract

Breslow thickness (BT) is the most powerful prognostic indicator in cutaneous melanoma, yet histopathological measurement exhibits some limitations such as interobserver variability and diagnostic delays. Preoperative clinical assessment demonstrates 30% misclassification rates. This narrative review synthesizes evidence on deep learning models for non-invasive BT estimation from dermoscopic images. Convolutional neural networks (ResNet, EfficientNet, Vision Transformers) with transfer learning from ImageNet achieve up to 75–79% accuracy and AUC 0.76–0.85 on single-center datasets. Preprocessing techniques (hair removal, color normalization, data augmentation) and interpretability methods (Grad-CAM, LIME) enhance clinical applicability. However, external validation reveals performance degradation. The clinically critical thickness range (0.4–1.0 mm) demonstrates poor discrimination. Significant dataset bias exists: most training data represents lighter skin phototypes, resulting in an underrepresentation of darker skin types. AI models function as complementary decision-support tools rather than replacements for histopathology. Prospective clinical trials validating clinical utility are lacking, and regulatory approval pathways are undefined. Research priorities include diverse public datasets with balanced skin tone representation, the adoption of threshold-weighted loss functions to prioritize accuracy at the 0.8 mm surgical cut-off, multi-institutional external validation, prospective randomized trials, federated learning frameworks, and regulatory engagement. Only rigorous, equitable research can translate AI from proof-of-concept to clinically reliable tools benefiting all melanoma patients.