Machine learning approaches for skin cancer detection: A study based on the ISIC 2024 competition

Details

Supervisors

Pecheur, Charles

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThis master’s thesis investigates deep learning approaches for skin cancer detection from clinical images within the context of the ISIC 2024 challenge. It analyses the impact of architectural design choices, attention mechanisms, and pretraining strategies on model performance, robustness, and interpretability. The study examines multimodal integration by combining image-based predictions with clinical metadata and compares neural fusion methods with gradient boosting approaches. Experimental results show that multimodal fusion through gradient boosting significantly outperforms neural multimodal integration strategies. In addition, external validation on a dermoscopic dataset reveals that incorporating Triplet Attention improves cross-domain generalization from non-dermoscopic to dermoscopic images, despite similar internal cross-validation results across model variants. Grad-CAM analysis further highlights differences in model attention induced by different training strategies.