Towards Generalizable Multi-sequence MRI Analysis: From Specialist to Foundation Models

PhD Thesis Proposal Defence


Title: "Towards Generalizable Multi-sequence MRI Analysis: From Specialist to
Foundation Models"

by

Mr. Zelin QIU


Abstract:

Multi-sequence magnetic resonance imaging (MRI) is a cornerstone of clinical 
diagnosis. Different MRI sequences highlight distinct anatomical and 
pathological features, providing critical information that assists 
radiologists in characterizing tissue properties, identifying lesions, and 
comprehensively assessing pathological changes. However, the wide variability 
of scanning protocols, including differences in magnetic field strength, 
pulse sequences, and imaging parameters, introduces substantial inter- and 
intra-sequence heterogeneity. Such heterogeneity poses significant challenges 
to deep learning models, limits their generalizability, and thus necessitates 
robust, adaptable approaches for diverse clinical settings.

This thesis addresses these challenges through three sequential studies. 
First, we develop CoPAS, demonstrating that explicitly modeling the 
relationship between imaging planes and pathology yields accurate and 
interpretable diagnoses for twelve knee abnormalities. This specialist model 
exhibits plane preferences consistent with independent clinical knowledge. 
Second, we present KAMRNet, extending this line of work to a larger data 
scale and comprehensive clinical validation. The model improves radiologists' 
accuracy and reduces reading time while also revealing the annotation cost 
that constrains fully supervised learning. Third, we introduce MARS, which 
bypasses this bottleneck through pretraining on large-scale multi-sequence 
MRI scans across ten anatomical regions. By learning disentangled 
representations across MRI sequences, MARS transfers effectively to versatile 
downstream tasks under heterogeneous data settings.

Together, these studies trace a progression from an interpretable specialist 
model through rigorous clinical validation to a generalizable foundation 
model, establishing a feasible pathway toward robust, clinically applicable 
multi-sequence MRI analysis.


Date:                   Monday, 5 October 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Dr. Hao Chen (Supervisor)
                        Dr. Long Chen (Chairperson)
                        Dr. Sehi L'Yi