A Survey on Explainability in Biomedical Vision-Language Foundation Models

PhD Qualifying Examination


Title: "A Survey on Explainability in Biomedical Vision-Language
Foundation Models"

by

Mr. Yuxiang NIE


Abstract:

Biomedical vision-language foundation models learn joint representations from 
medical images and clinical text, enabling tasks such as zero-shot 
classification and report generation. However, their decision processes remain 
largely opaque, making it difficult to determine whether predictions depend on 
clinically meaningful evidence or on statistical shortcuts. In this survey, we 
review explainability methods for biomedical vision-language models, 
organizing them into four perspectives: post-hoc attribution and 
representation analysis, visual-textual grounding, concept-based 
interpretation and prediction, and natural language explanations from 
generative models. Post-hoc methods estimate which input features influence a 
model's output. Grounding methods connect clinical phrases to specific image 
regions. Concept-based methods route predictions through named medical 
attributes, making the decision structure transparent. Generative methods 
produce natural language rationales that resemble clinical reasoning. We 
discuss open challenges, including incomplete concept vocabularies, 
hallucination in generated text, and limited generalization across clinical 
settings. Finally, we outline research directions toward more verifiable and 
clinically useful model explanations.


Date:                   Monday, 26 October 2026

Time:                   4:00pm - 5:30pm

Venue:                  Room 5501
                        Lifts 25/26

Committee Members:      Dr. Hao Chen (Supervisor)
                        Dr. Dan Xu (Chairperson)
                        Dr. Terence Tsz Wai Wong (CBE)