MICS 研讨会：Théo Dipiazza

Abstract
Computed Tomography (CT) is a key modality in medical imaging, but the growing volume of examinations highlights the need for automated tools to assist radiologists. Developing such tools for 3D CT remains challenging due to the complexity of volumetric data, limited annotations, and the diversity of abnormalities encountered in clinical practice. This research explores automated understanding of 3D CT scans, with a particular focus on abnormality classification and report generation. An abnormality-guided approach is first investigated, using intermediate abnormality predictions to improve the clinical relevance of generated reports. This is followed by the development of 2.5D representations based on interacting axial slices, providing an efficient alternative to fully 3D architectures for volumetric CT analysis. These ideas are further extended to both global and fine-grained slice-level vision-language pre-training, enabling the alignment of CT volumes with radiological reports. Finally, unified multi-task learning strategies are explored to jointly address complementary aspects of CT understanding, including abnormality localization, classification, and report generation. Overall, this work investigates how efficient volumetric representations, multi-modal learning, and task sharing can contribute to more robust and clinically relevant AI systems for 3D CT imaging.
Biography
Theo Di Piazza is a PhD student at INSA Lyon and Hospices Civil de Lyon, supervised by Prof. Loic Boussel. His research focuses on automated report generation from CT volumes, with an emphasis on 2.5D modeling and applications to abnormality classification. Homepage: https://theodpzz.github.io/.