Speaker
Description
Skin cancer is among the most prevalent cancers worldwide, and current monitoring relies on biweekly image-based follow-ups that visually compare lesion changes over time. Although effective, this strategy is reactive and offers limited ability to anticipate future lesion behavior. To address this gap, we propose a predictive framework that leverages routinely collected clinical images to forecast lesion evolution. The approach combines a convolutional autoencoder, which encodes lesion images into a compact latent representation capturing key visual features, with a machine learning model that models temporal changes within this latent space. The decoder reconstructs images from latent vectors, enabling generation of future lesion appearances. Analysis of the latent space revealed structured, interpretable representations of lesion dynamics. From these, quantitative metrics such as trajectory length, progression speed, and cluster transition frequency were derived. Notably, these descriptors effectively differentiated stable lesions from rapidly evolving ones and captured patient-specific progression patterns. Lesions with higher latent speeds and more frequent cluster transitions showed greater morphological change over time. Overall, this framework provides a quantitative basis for anticipating lesion progression, supporting more proactive and personalized treatment decisions, including therapy initiation, adjustment, and response prediction.
