Principal Investigator: Debora Gil
Email: debora.gil@cvc.uab.cat
Web: http://iam.cvc.uab.es/; https://www.cvc.uab.es/research-lines/interactive-and-augmented-modelling/
Brief Theme Description:
Under the starting hypothesis that tumoral pathological heterogeneity is key for the definition of personalized therapy, this research focuses on clinical meaningful representation learning mechanisms able to describe such pathological contexts. We will apply these mechanisms to the characterization of tumour heterogeneity and genomic profiles as new visual biomarkers in cancer early diagnosis and personalized therapy.
The specific problems linked to Health addressed in the research will include on-going projects in tight cooperation with clinical entities:
- Tumor tissue heterogeneity in Computer Tomography (CT) scans as new radiogenomic signature in lung cancer for personalized treatment (Hospital Germans Trias i Pujol).
- Tumor tissue heterogeneity in histological Whole Slide Images (WSI) as new pathomic biomarker of the prediction of risk of lymph node metastasis and survival rate in early-stage cancer (Institut Josep Carreras, Hospital Clinic of Barcelona).
Tissue structure will aggregate local features of WSI/Medical Scan visual appearance to obtain a hierarchical representation of the global content using learnable embeddings of visual and topological descriptors based on persistent homology. The candidate will use foundational models to extract features from a collection of tissue image patches and will have to develop several aggregation strategies, such as hierarchical attention and GATs with differentiable poolings.
The hierarchical embeddings will be learned for two different models: a target-oriented one trained to predict clinically relevant pathological aspect (risk of metastasis, aggressiveness, response to treatment); a clinically-driven one trained to map target-oriented embeddings into clinical words to provide a clinical interpretation. For increased explainability, clinically-driven and target-oriented attention weights will be compared in a cross-explainability module to identify which components contribute to, both, prediction of diagnosis/prognosis and description of the lesion in pathological/radiological terms.
Clinical information will also be included using LLMs and hierarchical attention mechanisms for its integration with the visual information.
Keywords: Hierarchical Representation Learning, Persistent Homology, Multimodal Personalized Medicine; Deep Topological Learning; Explainable and Fair ML
