More about HKUST
Generalizable AI for Computed Tomography: From Volumetric Pretraining to Clinical Validation
PhD Thesis Proposal Defence
Title: "Generalizable AI for Computed Tomography: From Volumetric Pretraining
to Clinical Validation"
by
Mr. Jiaxin ZHUANG
Abstract:
Computed tomography (CT) is a widely used volumetric imaging modality with
broad clinical applications, including disease detection, diagnosis, staging,
treatment planning, and longitudinal assessment. Recent CT foundation models
have enabled a broader range of downstream CT analysis tasks and achieved
strong performance across retrospective benchmarks, yet their generalizability
and clinical utility remain insufficiently established, especially in
prospective and human-AI settings. Developing generalizable CT AI remains
challenging because CT volumes are high-dimensional, contain complex
anatomical and a broad range of abnormal findings. Beyond effective volumetric
learning, clinically relevant CT AI also requires clinical semantic grounding
and rigorous evaluation across diverse clinical settings.
This thesis addresses these challenges through four sequential studies.
Bio2Vol shows that knowledge from pretrained two-dimensional biomedical
foundation models can be effectively transferred to volumetric medical
imaging. GL-MAE develops dedicated 3D self-supervised pretraining that jointly
models global anatomical context and local structural detail, yielding
improved transfer performance across downstream volumetric tasks. MiM further
introduces hierarchical pretraining across spatial scales and demonstrates
additional gains with increasing pretraining data. CARE extends these advances
to large-scale CT-report vision-language pretraining, demonstrating broad
transfer across heterogeneous tasks and data settings, followed by external,
prospective, and human-AI validation.
Together, these studies establish a progression from leveraging existing
biomedical knowledge, through dedicated and scalable volumetric pretraining,
to vision-language foundation modeling and clinical validation. The findings
indicate that generalizable CT AI requires coordinated advances in volumetric
pretraining, clinical semantic grounding, broad transfer, and rigorous
clinical validation.
Date: Monday, 28 September 2026
Time: 10:00am - 12:00noon
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Hao Chen (Supervisor)
Dr. Long Chen (Chairperson)
Dr. Sehi L'Yi