Yazan-Aref/tb-crispr-cxr-demo

TB Chest X-ray Classifier — TB-CRISPR research demo

A DenseNet121 fine-tuned on chest X-rays for binary tuberculosis classification. Upload a chest X-ray (PA view); the demo returns:

• —Prediction: TB-suggestive vs. Normal/No TB pattern
• —Probabilities: P(TB) and P(Normal) from the sigmoid head
• —Grad-CAM heatmap: highlights which regions of the X-ray the model relied on

⚠️ Disclaimer

This is a research demo. It is NOT a medical device, NOT clinically validated, and MUST NOT be used for diagnosis, screening, or any clinical decision. Always consult a qualified clinician. Confirmatory testing (culture, Xpert MTB/RIF, or CRISPR-Cas) is required.

Project

This demo is the visual triage component of the **TB-CRISPR / IAS-MD** diagnostic pipeline (ITEX 2026): a CRISPR-Cas TB detection assay paired with AI image triage and a lyophilized field-deployable format. In the full pipeline this model only flags candidates; confirmation comes from molecular detection.

Model details

• —Architecture: torchvision DenseNet121
• —Head: single-logit binary classifier (Linear(1024, 1) → sigmoid)
• —Input: 224×224 RGB, ImageNet mean/std normalization
• —Grad-CAM target layer: features.norm5

Public CXR datasets for testing

• —Montgomery & Shenzhen TB sets — NIH
• —TBX11K (Kaggle)

License

Code: CC-BY-NC-4.0 (non-commercial research use). Model weights: research use only, not for clinical deployment.