All projects

Research · Final Year Project · B.Sc. Computer Engineering

MedVLM-Report

Final Year Project — IEEE Conference Paper

evaluation-summary.log
$ python evaluate.py --config medvlm_report.yaml
[Stage-1] DenseNet ensemble (x5) — chest X-ray findings
[Stage-1] ResNet-18 (weak-sup) — breast ultrasound findings
[Backbone] Swin + BLIP-2 Q-Former -> LoRA-adapted BioGPT decoder
[Loss] Faithfulness-weighted detect-then-describe conditioning
Held-out studies evaluated: 1,703
Mean detector AUC: 0.704 | Clinical F1: 0.451
Explainability: Grad-CAM, attention rollout, sentence-to-region ✓

A unified vision-language framework that generates radiology reports for both chest X-ray and breast ultrasound from one shared decoder. The central contribution is a detect-then-describe conditioning mechanism: a Stage-1 finding detector's output is injected into the language decoder's prompt and constrained with a faithfulness-weighted loss, rather than relying on implicit visual grounding. Every report ships with a four-part explainability package — attention rollout, Grad-CAM, sentence-to-region grounding, and token/sentence rationale — evaluated across a 14-version ablation on 1,703 held-out studies.

Stack

  • Vision-Language Models
  • PyTorch
  • BioGPT
  • LoRA
  • Swin Transformer
  • Q-Former
  • Explainable AI

Next project

KCSC — Kallar Central Sports Club