Are deep features able to learn melanoma depth gradation when trained only for binary tasks?

This study aims to analyze dermoscopic images to estimate the Breslow depth of melanoma, which is crucial for diagnosing and treating skin cancer. The study has two objectives: to improve current methods of estimating melanoma depth and to investigate whether dermoscopic features evolve as melanoma thickness increases. To achieve these goals, a Convolutional Neural Network (CNN), with ConvNext as backbone, was designed and trained to predict melanoma thickness. Then, the relationships between the CNN’s outputs and the continuous depth values from biopsies were explored to investigate possible gradations in deep features related to melanoma depth. First, statistical analyses were performed in the deep feature spaces. Secondly, deep features were transformed into Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) spaces. Two main results can be reported. First, the model outperformed existing methods in the literature for classifying melanomas as superficial or deep. It achieved the highest recall (0.79) and F1 score (0.67), a good balance between precision and recall. As for the relationships between deep features and biopsy depth, a moderate correlation was found between melanoma depth and prediction probabilities, which was stronger in deeper melanomas. Both PCA and UMAP analyses revealed that intermediate-depth melanomas have deep features located in intermediate positions in the transformed feature space. These findings highlight the capacity of deep features to distinguish between different depths of melanoma. The study provides new insights into the data distribution and suggests a potential methodology for enhancing the understanding of melanoma classification through deep learning models. • Optimization strategies of Deep learning models (ConvNext, RegNet) can improve binary melanoma thickness prediction classification, up to 0.79 recall. • CNNs, trained for binary tasks, still capture gradations in melanoma thickness as depth increases. • PCA and UMAP feature transformations reveal a clear distinction between thin and thick melanomas, with intermediate cases showing mixed characteristics. • Explanability based on analysis of deep feature distributions, rather than activation maps, provides insight into how CNNs learn melanoma depth even without explicit thickness labels.