| Resource | Link |
|---|---|
| Demo Video | https://www.youtube.com/watch?v=1Haj5EFSahw |
| Presentation Link | https://gamma.app/docs/PROBLEM-STATEMENT-13dzdo0ba29p1w5?mode=present#card-rpp0rb40e5f4lfr |
TB-Vision is a clinical decision support system for tuberculosis screening that combines lightweight deep learning models with explainable AI and intelligent validation. The system is designed for rural clinics and low-resource healthcare environments, where radiologists and diagnostic infrastructure are limited.
1️⃣ Local CNN ensemble analyzes chest X-rays
2️⃣ Uncertainty estimation determines prediction confidence
3️⃣ Cloud validation is triggered for further medical explaination
This hybrid design enables fast, affordable, and scalable TB screening worldwide.
Most AI systems for medical imaging suffer from:
- black-box predictions
- overconfident outputs
- lack of clinical context
- dependence on cloud infrastructure
TB-Vision solves these problems through:
- Explainable AI (Grad-CAM++)
- Uncertainty-aware predictions
- Offline-first deployment
- Multi-stage AI validation
Tuberculosis remains one of the deadliest infectious diseases worldwide.
- 10.7 million cases reported in 2024
- 1.23 million deaths annually
- 2.4 million cases remain undiagnosed
Many countries with the highest TB burden lack access to diagnostic radiology.
| Region | Radiologists per million |
|---|---|
| USA / Europe | 100+ |
| Indonesia | <10 |
| Pakistan | <8 |
| Low-income regions | <2 |
Over 50% of the world's population lacks reliable diagnostic imaging access.
Existing AI tools often fail in real clinical environments because they:
- act as black boxes
- produce overconfident predictions
- require constant internet connectivity
- are too expensive for mass screening
This creates a critical need for an affordable, explainable, and offline-capable TB screening system.
TB-Vision introduces a hybrid AI screening system that combines:
- lightweight deep learning models
- explainable AI
- uncertainty-aware predictions
- optional cloud validation
Offline First
The core CNN ensemble runs locally on basic computers without internet access.
Explainability
Grad-CAM++ highlights the lung regions influencing the AI decision.
Uncertainty Awareness
Monte-Carlo Dropout estimates prediction confidence and flags risky cases.
Intelligent Escalation
Only uncertain cases are forwarded to advanced AI models for deeper analysis.
1️⃣ Patient X-ray uploaded
2️⃣ CNN ensemble performs local prediction
3️⃣ Uncertainty score calculated
4️⃣ High-confidence cases resolved locally
5️⃣ Uncertain cases escalated for AI validation
TB-Vision follows a multi-stage AI pipeline.
Models used:
- DenseNet121
- EfficientNet-B3
- ResNet50
The ensemble improves robustness and reduces model bias.
Outputs:
- TB probability
- prediction uncertainty
- Grad-CAM heatmap
Monte-Carlo Dropout performs multiple forward passes to measure prediction confidence.
This helps detect cases where the model may be unsure.
When uncertainty is high, the system can optionally use cloud AI models to validate findings and generate clinical explanations.
This ensures safety without requiring constant internet connectivity.
- AUC: 0.9978 – near-perfect discrimination between TB and non-TB cases.
- F1: 0.9781 – maintains strong balance between precision and recall.
- Accuracy: 0.9769 – reflects consistent predictions on the validation set.
- Sensitivity: 0.9787 and Specificity: 0.9883 – demonstrates both high true positive and true negative rates.
- Class accuracy: [Normal: 0.984, TB: 0.956, Other pathologies: 0.996].
- System architecture: the full TB-Vision pipeline from CXR upload through explainability and escalation.
- Classifier decisions: ensemble output with probabilities, uncertainty, and Grad-CAM overlays situating the model rationale.
- Confusion matrix: per-class prediction breakdown, highlighting the ensemble’s precision and recall trade-offs.
- ROC curve: discrimination capability across thresholds for the TB versus non-TB classification.
