All projects
ResearchJan – Aug 2025Computer Vision

Multi-Modal Vision-Language Model for Cancer Diagnosis

Cross-modal pipeline correlating high-dimensional medical imaging with unstructured biomedical literature, with a Grad-CAM + Gradio explainability interface for clinical interpretability.

Pipeline

MRI imaging
Vision Transformer
BioBERT fusion
Grad-CAM
Gradio UI

The problem

Tumour assessment reads imaging and clinical narrative separately, and single-modality deep models return a label without showing which region of the scan drove it — which clinicians cannot act on.

Architecture

End-to-end flow, stage by stage.

  1. 01

    MRI imaging

  2. 02

    Vision Transformer

  3. 03

    BioBERT fusion

  4. 04

    Grad-CAM

  5. 05

    Gradio UI

How it works

01

Single-modality baselines first

Separate notebooks establish reference points: BioBERT and ClinicalBERT on the clinical report corpus, and a Vision Transformer on the MRI images, so cross-modal gains can be attributed rather than assumed.

02

Cross-modal fusion

Image features and clinical text representations — including synthetic text experiments — are combined in a cross-modal training notebook, with dedicated inference and evaluation notebooks reporting accuracy, F1 and AUROC on cleaned train/validation splits.

03

Explainability in the loop

Grad-CAM heatmaps are surfaced through a Gradio interface so a reviewer can see the image regions behind each prediction instead of trusting a bare score, reaching 80% validation accuracy on the MRI tumour task.

04

Reproducible data contract

Raw and cleaned CSVs are kept side by side with an explicit input/output contract, so notebook runs are repeatable and the pipeline extends to other cancer types, modalities and text sources.

  • Fused Vision Transformer features from MRI scans with BioBERT and Clinical BERT representations of clinical text to support cross-modal diagnostic reasoning.
  • Reached 80% validation accuracy on MRI tumor data and exposed Grad-CAM heatmaps through a Gradio interface so clinicians could inspect the image regions influencing each prediction.
  • Designed the inference pipeline to extend across additional cancer types, imaging modalities, and biomedical text sources rather than remain tied to one dataset.

By the numbers

From text-only to explainable cross-modal

Validation accuracy (%)

higher is better
Value
80%

Modalities used

higher is better
Before
1
After
2

Model families benchmarked

Value
4

What it ships with

  • BioBERT and ClinicalBERT text baselines
  • Vision Transformer imaging baseline
  • Cross-modal training and inference notebooks
  • Grad-CAM explainability via Gradio
  • Cleaned, versioned train/validation metadata
ImpactBefore → After
Text-only reports80%
Validation accuracy on MRI data
Black-boxGrad-CAM
Per-prediction explainability

Stack

PyTorchGrad-CAMGradioVLM