LETITIA

Motivation

‍Despite the transformative success of immune checkpoint inhibitors (ICIs) in oncology, treatment outcomes remain highly variable and difficult to predict, with a critical lack of clinically validated biomarkers to guide patient stratification. Understanding distinct response patterns, such as dissociated response or pseudoprogression, remains a pressing clinical challenge, especially in cancers like melanoma where current imaging studies are limited and methodologically inconsistent. This project aims to address these gaps by developing advanced, data-driven tools that leverage longitudinal PET/CT imaging and radiomics to capture the spatial and temporal evolution of individual tumor lesions. By automating lesion tracking and integrating imaging biomarkers with clinical data, the goal is to build robust models that predict ICI response at an early stage, ultimately enabling more personalized and effective treatment strategies.

Proposed Approach / Solution

‍We propose to address these objectives by casting the problem of predicting ICI response in a spatiotemporal context, through two main stages. In the first stage, we will develop a longitudinal PET/CT segmentation and tracking algorithm that captures the temporal evolution of individual lesions and uses this information to improve segmentation accuracy and lesion detection. This will be achieved by building on recent advances in medical foundation models trained across diverse imaging modalities. In the second stage, we will design a dynamic graph representation learning framework that models both intra- and inter-lesion dependencies. The lesion trajectories extracted in the first stage will serve to combine PET/CT radiomics features from a patient’s metastatic lesions, clinical data such as blood biomarkers, and informative representations from foundation model embeddings, to predict ICI treatment outcomes for melanoma and lung cancer patients. These methods will first be developed and validated on the USZ-Melanoma dataset, which includes ground-truth lesion detection and tracking. The CHUV-Melanoma and CHUV-Lung cancer datasets, which contain noisier annotations, will then be used to evaluate the generalizability of both frameworks.

‍

‍

Impact

‍The proposed research will demonstrate the power of data science in personalized oncology by developing an automated algorithm for longitudinal segmentation and lesion tracking from routine clinical PET/CT scans. This technology will enable large-scale, automated curation of real-world imaging data, supporting research into tumor dynamics and offering potential clinical applications in monitoring treatment response. By identifying biomarkers of ICI response from non-invasive imaging, the project aims to improve patient-specific treatment stratification, optimizing outcomes while minimizing adverse effects.