Medical Vision-Language Models
Understanding how multimodal foundation models represent and reason about radiology images and clinical concepts.
Nafiz Sadman
PhD Researcher in Medical Vision-Language Models
I study how multimodal AI systems understand medical images, particularly when data are imbalanced, findings are rare, or model behaviour is difficult to interpret.
Research focus
My current research explorations
Medical images are complex, rare findings matter, and reliable model behaviour needs careful evaluation.
Long-Tailed Medical AI
Improving recognition and evaluation for rare and underrepresented findings in real-world medical datasets.
Interpretable & Reliable AI
Studying model behaviour, robustness, grounding, calibration, representation, and explanation quality.
Featured research
Projects in focus.
Research into evaluation, imbalance, and interpretation in medical imaging.
Chest CT · VLM evaluation
CTVLM-Inspect
A model-agnostic evaluation framework for chest-CT vision-language models spanning classification, calibration, retrieval, reconstruction consistency, interpretability, and related model behaviour.
Explore research →
Rare findings · Radiology
Long-Tailed Learning in Chest Radiology
A research program on rare-finding recognition, cross-dataset behaviour, calibration, grounding, and adaptation strategies for chest X-ray and CT models.
Explore research →
BioKDD 2025
Biomedical VLM Interpretability
Understanding representation and explanation behaviour under class imbalance and distribution shift.
Read publication →
IEEE ICC 2026
DepthPulse+
Combining depth information and physiological signals for face presentation-attack detection.
Read publication →Selected publications
A few representative papers.
DepthPulse+: A Depth and Vital Sign Based Method for Face Presentation Attack Detection
IEEE ICC 2026
Presentation-attack detection using depth and physiological cues.
Interpreting Biomedical VLMs on High-Imbalance Out-of-Distributions
BioKDD 2025
Interpretability analysis of biomedical VLM behaviour in radiology settings.
Visualizing Temporal Interval Hierarchies
NASA Formal Methods 2025
Visualization tooling for temporal interval structures in streaming system data.
Diagnosis and management of Alzheimer's disease in primary care
Journal of Primary Care & Community Health
A real-world Ontario clinical study informed by electronic health-record analysis.
Research internships & collaborations
Research beyond the lab.
Work with EvenUp, Pfizer Canada, and Queen's collaborators connects machine learning to legal retrieval, clinical research, and visualization.
EvenUp
Associate Machine Learning Engineer · Internship
–
Retrieval & ranking for legal AI
Built relevance evaluation and reranking methods for legal information retrieval.
- LLM-as-judge
- Cross-encoder, listwise & setwise reranking
- Precision@k · Recall@k · NDCG@k
Pfizer Canada
ML Research Intern · MITACS Accelerate
–
Real-world clinical data analysis
Analyzed unstructured electronic health records to study patient characteristics and comorbidity patterns.
- EHR transformation
- Clustering
- Logistic regression
Academic collaborations
CritLab@Queen's ECE
Manager: Dr. Sean Kauffman
Co-created a visualization interface for space log analysis and nfer, bringing streaming events and temporal interval hierarchies into interactive views. Built the interface with Python Socket.IO and D3.js.
Python Socket.IO · D3.js · Streaming visualization
Faculty of Health Science, Queen's University
Supervisors: Prof. Farhana Zulkernine and Dr. Nicholas Cofie
Automated the extraction and evaluation of resident Entrustable Professional Activities from assessor notes. Analyzed sentiment and Spearman correlation between assessor feedback and committee comments.
School of Rehabilitation Therapy, Queen's University
Supervisors: Prof. Farhana Zulkernine and Assoc. Prof. Nora Fayed
Automated literature filtration, information extraction, and classification across 8,000 clinical randomized controlled trial articles using domain-specific BERT, with a reported 95% specificity. Contributed to grant writing.