Generative AI-Enabled Virtual Dynamic Contrast-Enhanced MRI for Breast Cancer Diagnosis

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


MPhil Thesis Defence


Title: "Generative AI-Enabled Virtual Dynamic Contrast-Enhanced MRI for
Breast Cancer Diagnosis"

By

Miss Yi XIN


Abstract:

Gadolinium-free breast MRI represents a promising direction for safer and 
more accessible breast cancer diagnosis. The foundation of this goal lies in 
preserving the diagnostic enhancement information of dynamic 
contrast-enhanced MRI (DCE-MRI) without administering gadolinium-based 
contrast agents. Despite growing interest in non-contrast and 
contrast-sparing breast MRI, existing approaches often lose 
enhancement-related diagnostic information or lack rigorous clinical 
validation. This thesis presents vDCE, a diagnostic-aware generative AI 
framework for virtual DCE-MRI synthesis, addressing the unmet need for 
contrast-free enhancement imaging in breast MRI. Unlike conventional image 
translation methods that primarily optimize visual similarity, vDCE uses 
multimodal non-contrast MRI and a dual-task bridge diffusion model to jointly 
synthesize virtual DCE-MRI and T1-subtraction images, encouraging 
preservation of contrast-induced lesion information. Using multicentre 
retrospective cohorts, a prospective cohort, benchmark comparisons, AI-based 
diagnostic evaluation, and a blinded multi-reader crossover study, vDCE 
generated high-fidelity virtual DCE-MRI, achieving an internal-test SSIM of 
0.808 and PSNR of 27.66 dB. Virtual DCE-MRI improved malignancy 
classification over non-contrast MRI alone, increasing AUROC from 0.707 to 
0.808 in the internal test cohort, from 0.762 to 0.926 in an external cohort, 
and from 0.543 to 0.914 in the prospective cohort. In the reader study, 
radiologists achieved non-inferior diagnostic performance using virtual 
DCE-MRI compared with real DCE-MRI, with AUROC values of 0.954 and 0.946, 
respectively, and substantial BI-RADS agreement (kappa=0.742). These findings 
highlight the potential of diagnostic-aware virtual DCE-MRI to reduce 
gadolinium dependence and advance safer, more accessible breast MRI 
workflows.


Date:                   Monday, 27 July 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Dr. Terence Tsz Wai WONG (CBE)

Committee Members:      Dr. Hao CHEN (Supervisor)
                        Prof. Can YANG (MATH)