Tuberculosis Detection from Chest X-ray using CNN

Tuberculosis remains one of the world's deadliest infectious diseases, and chest X-ray screening is a frontline triage tool in high-burden regions — but there are far fewer trained radiologists than X-ray machines. Computer-aided screening that flags TB-suggestive patterns (infiltrates, cavities, upper-lobe consolidation) can help prioritize which films a radiologist reads first. This project builds that screening idea as an honest educational prototype: a custom CNN trained from scratch on the two classic public TB X-ray collections from the U.S. National Library of Medicine, evaluated with ROC-AUC, sensitivity and specificity rather than a single accuracy number. The interactive demo lets you select real sample radiographs, run the analysis, and inspect where the network attended — while every page of the project states plainly that this is a teaching build, not a diagnostic device.

1. The Montgomery (138) and Shenzhen (662) chest X-ray sets are combined — 800 frontal CXRs labeled normal vs TB — resized to 224×224 grayscale, histogram-equalized and normalized.
2. Stratified train/validation/test splits preserve the class ratio; augmentation (small rotations, flips, contrast jitter) is applied during training.
3. The custom CNN — 4 conv blocks → global average pooling → Dense(128, dropout 0.5) → sigmoid — is compiled with Adam (lr 1e-4) and weighted binary cross-entropy to handle class imbalance.
4. Training runs with early stopping on validation AUC; the best checkpoint is kept and the curves recorded.
5. The held-out test set is evaluated once: ROC-AUC, sensitivity, specificity and the confusion matrix are generated for the report.
6. In the web demo, a selected radiograph is preprocessed identically and passed through the saved weights; the app shows probability bars and the attention overlay. All outputs are labeled illustrative — not a diagnosis.

Tech stack:

• Python 3, TensorFlow/Keras
• NumPy, Matplotlib, scikit-learn (metrics, ROC)
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (interactive radiograph demo)
• Montgomery + Shenzhen TB X-ray datasets (NLM, 800 CXRs)

Dataset & model details

• Dataset: Montgomery County X-ray Set (U.S. National Library of Medicine) — 138 frontal chest X-rays (80 normal, 58 TB); Shenzhen Hospital X-ray Set (U.S. National Library of Medicine) — 662 frontal chest X-rays (326 normal, 336 TB). Combined: 800 CXRs, binary labels (normal / TB).
• Task: Binary image classification; input = 224×224×1 grayscale radiograph, output = probability of TB-suggestive pattern.
• Model: Custom CNN (~2.1M parameters, design target): 4× (Conv 3×3 → BatchNorm → ReLU → MaxPool) with 32/64/128/256 filters → global average pooling → Dense 128 (dropout 0.5) → sigmoid. Weighted binary cross-entropy, Adam (lr 1e-4), early stopping on validation AUC.
• Metrics: Validation ROC-AUC ≈ 0.90, sensitivity ≈ 0.85, specificity ≈ 0.88 (design targets for the built-to-order training run), confusion matrix, PR curve. No figure is claimed as measured until the training run is executed for the order.

• [Custom CNN for radiographs] Four convolution blocks (32/64/128/256 filters) with batch normalization and max-pooling on 224×224 grayscale input, global average pooling, Dense(128) with dropout 0.5 and sigmoid output — ~2.1M parameters, designed from scratch with a documented rationale.
• [Interactive radiograph analysis] Select from 5 real sample chest X-rays, run the CNN analysis, and see TB-suggestive vs normal/clear probability bars with an explanatory note per image.
• [Grad-CAM-style attention overlay] A schematic heatmap marks the lung zones the model weighted most for the selected scan — the standard explainability visual for medical imaging, labeled as illustrative in the demo.
• [Real public TB X-ray collections] Trained on the Montgomery County (138 CXRs) and Shenzhen Hospital (662 CXRs) sets from the U.S. National Library of Medicine — real, citable, public data.
• [Clinically honest evaluation] ROC-AUC, sensitivity, specificity and confusion matrix instead of a lone accuracy figure, with the class imbalance handled by weighted loss — the way medical imaging is actually evaluated.
• [Full training notebook] Preprocessing (histogram equalization, normalization), augmentation, model definition, training with early stopping on validation AUC, and evaluation in one reproducible notebook.
• [Training curves] Validation AUC and loss curves across epochs with interpretation notes, included in the report.
• [Exported trained model] Saved weights plus the preprocessing pipeline so the demo runs the real network without retraining.

• Complete training & evaluation Jupyter notebook
• Trained CNN weights with preprocessing code
• Interactive radiograph-analysis web demo with attention overlay
• ROC/PR curves, confusion matrix and training-curve plots
• Project report PDF (background, CNN theory, architecture rationale, methodology, results, medical-AI limits)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, class imbalance, AUC vs accuracy, Grad-CAM, why this is not a diagnostic device)

• An educational screening prototype — not a medical diagnostic device; no clinical claim is made and it must never guide patient care.
• 800 images is small for deep learning; the report discusses overfitting risk and the augmentation/regularization used to manage it.
• Frontal (PA/AP) views only — lateral views and CT are out of scope.
• TB labels in the public sets are image-level; the model does not localize lesions or distinguish active from latent TB.
• All performance figures are design targets for the training run, stated honestly — the report documents the actual achieved figures after training.