Tuberculosis Detection using Chest X-Rays

Tuberculosis still ranks among the top infectious killers, and the chest X-ray is the frontline screening tool where sputum testing is slow or unavailable. The radiographic signs — apical nodules, cavitation, fibrotic streaking — are recognizable but reading thousands of screening films needs trained radiologists who are scarce in high-burden regions. Two public datasets exist precisely for this research, released by the U.S. National Library of Medicine: the Shenzhen set (662 posterior-anterior chest X-rays, 336 TB and 326 normal) and the Montgomery County set (138 films, 58 TB and 80 normal). This project fine-tunes a DenseNet121 on these films, but its real distinguishing feature is methodological: instead of reporting only in-distribution scores, it trains on one institution and tests on the other, confronting the domain-shift problem that decides whether any screening model works in practice. The demo makes this visible with a dedicated cross-dataset comparison view, and the documentation carries the medical-prototype disclaimer throughout.

1. Chest X-rays from the Shenzhen and Montgomery sets are loaded and passed through a U-Net lung segmenter; each film is cropped to its lung ROI.
2. ROIs are contrast-enhanced with CLAHE, resized to 512x512 and normalized with ImageNet statistics.
3. Splits are made at the patient level; the cross-institution protocol additionally holds out one entire dataset for testing.
4. A DenseNet121 backbone (ImageNet weights) with a dropout + sigmoid head trains with class-weighted binary cross-entropy under the AdamW optimizer, head first, then deeper blocks unfrozen.
5. Augmentation — small rotations, horizontal flips, elastic distortion — is applied during training; every epoch logs validation metrics and the best checkpoint is kept.
6. Evaluation reports AUC, sensitivity at 90% specificity and F1 both in-distribution and cross-institution.
7. The demo app serves predictions with attention overlays, the pattern checklist, and the side-by-side dataset comparison view.

Tech stack:

• Python 3.10, PyTorch with torchvision DenseNet121
• OpenCV (loading, CLAHE, ROI cropping), NumPy, pandas
• scikit-learn (patient-level splits, metrics)
• Matplotlib, Seaborn (training curves, ROC curves)
• Flask demo app (classifier, attention, dataset-comparison views)
• Shenzhen TB X-ray set (662 CXRs) + Montgomery County set (138 CXRs), via NLM

• [DenseNet121 transfer learning] ImageNet-pretrained DenseNet121 fine-tuned with a two-phase schedule (head first, then deeper dense blocks via AdamW) for TB-vs-normal classification
• [Lung-segmentation pre-step] U-Net lung masks crop each film to the lung ROI before classification, so the model learns from lung tissue rather than background and annotations
• [Cross-institution evaluation] Train-on-Shenzhen/test-on-Montgomery protocol (and reverse) with AUC, sensitivity-at-90%-specificity and F1 reported per direction — the honest generalization test
• [Attention overlays] Grad-CAM-style maps from the final dense block highlight the regions (typically upper lobes) driving each decision
• [Radiographic pattern checklist] Demo overlay annotating nodules, cavitation, fibrotic streaking, effusion and miliary pattern for teaching purposes
• [Three-view demo app] Classifier view, attention-map toggle, and a dataset-comparison view showing both collections side by side
• [512x512 high-resolution pipeline] CLAHE enhancement and ImageNet normalization tuned for the digitized film quality of these sets
• [Viva-ready documentation] Report with dataset analysis, domain-shift discussion, error analysis and the medical-prototype disclaimer stated explicitly

• Complete source code (segmentation, training, evaluation, demo app)
• Fine-tuned model weights exported from the included training run
• Jupyter training and evaluation notebook (buyer-run procedure: train, cross-test, evaluate)
• Project report PDF (background, dataset analysis, domain-shift discussion, evaluation, error analysis, medical-prototype disclaimer)
• PPT presentation for final review
• Viva Q&A preparation document (DenseNet, transfer learning, domain shift, metrics)
• Setup guide (environment, dataset download, training, running the demo)

• This is an academic research prototype, not a medical device: TB diagnosis requires sputum testing (e.g. CBNAAT) and a radiologist; this system never replaces that pathway.
• No AUC or sensitivity figure is promised: cross-institution scores are expected to be lower than in-distribution ones, and the report documents the measured gap honestly.
• The two datasets are small by deep-learning standards (800 films combined) and come from specific screening programs; performance on other populations is unvalidated.
• The model outputs a film-level score, not a localized diagnosis; it does not type TB manifestations (cavitation vs effusion) in the base build.
• Digitized-film artifacts and pediatric cases in the Shenzhen set add noise that bounds achievable performance.