Multimodal Learning for Biomedical Data Integration
Development of deep learning methodologies for integrating heterogeneous biomedical data, including radiological imaging, histopathology, and clinical text. Main applications include multimodal brain tumor segmentation and the automated analysis of whole-slide images and pathology reports.
Robust Multimodal Learning under Missing Modalities
Development of multimodal models designed to operate when one or more input modalities are unavailable during training and/or inference. The research covers vision-language and audio-video-language applications, as well as brain tumor segmentation from incomplete multimodal medical imaging.
Multi-Omics Integration
Development of deep learning methodologies for integrating spatial and non-spatial omic data in cancer research. The research aims to characterize cell-type composition and the spatial and molecular heterogeneity of tumor tissue while accounting for batch effects and technical variability across samples.
Explainable Artificial Intelligence for Multimodal Models
Development of methods for interpreting multimodal models, with a particular focus on tracing cross-modal information flow. The research investigates how multimodal large language models integrate and exploit visual and textual information throughout the generation process.
