A Deployable Explainable Deep Learning System for Tuberculosis Detection from Chest X-Rays in Resource-Constrained High-Burden Settings

This study presents and evaluates a deployable, explainable deep learning system based on DenseNet121 and Grad-CAM that achieves accurate tuberculosis detection from chest X-rays on both desktop and mobile platforms, demonstrating its potential as an offline decision support tool for resource-constrained healthcare settings.

The Gist

Imagine you are a doctor in a remote village. You have a patient with a cough, and you need to know if they have Tuberculosis (TB), a serious lung disease. In a perfect world, you'd have a team of expert radiologists and high-speed internet to send the X-ray for analysis. But in many parts of the world, you might only have a single X-ray machine, no internet, and no specialist nearby.

This paper introduces a solution called TBAI Africa: a smart, pocket-sized computer program that acts like a "second pair of eyes" for doctors in these tough situations.

Here is the story of how they built it, explained simply:

1. The Problem: The "Overworked Detective"

TB is a huge killer, especially in poorer countries. Catching it early is key. Doctors use chest X-rays to look for it, but reading an X-ray is hard. It's like trying to find a tiny needle in a haystack while wearing foggy glasses. If a doctor misses a case, the patient gets sicker. If they get it wrong, they might treat someone who isn't sick.

2. The Solution: A "Digital Intern" Who Never Sleeps

The researchers built an Artificial Intelligence (AI) system to help. Think of this AI as a super-intern who has studied millions of X-rays.

• The Brain: They didn't build the brain from scratch. Instead, they took a pre-trained brain (called DenseNet121) that was already good at recognizing things in pictures (like cats, cars, and trees) and taught it specifically to look at lungs. This is like taking a master chef who knows how to cook everything and teaching them specifically how to bake the perfect apple pie.
• The Training: They showed this AI thousands of X-rays, some from healthy people and some from people with TB. They taught it to spot the subtle differences, like a detective learning to spot a specific fingerprint.

3. The "Magic Glasses": Making the AI Honest

One big problem with AI is that it's a "black box." You give it a picture, it says "TB," but you don't know why. Doctors don't trust things they can't understand.

To fix this, the team added a feature called Grad-CAM.

• The Analogy: Imagine the AI is a student taking a test. Usually, you just see the final grade. Grad-CAM is like the teacher highlighting exactly which sentences in the textbook the student used to get the answer right.
• How it works: When the AI says, "This looks like TB," it also draws a glowing red heatmap over the X-ray, showing exactly which part of the lung made it think that. If the red glow is on the lung tissue, the doctor trusts it. If the red glow is on the edge of the photo or a metal clip, the doctor knows the AI is confused.

4. The "Offline" Superpower: No Internet Needed? No Problem!

Most fancy AI needs a super-fast internet connection to send data to a cloud server to get an answer. But in remote villages, the internet is slow or non-existent.

The researchers did something clever: they shrunk the giant AI brain down into a tiny, lightweight version called TensorFlow Lite.

• The Analogy: Imagine taking a massive library of books and compressing them into a single, tiny USB drive that fits in your pocket.
• The Result: They put this tiny AI onto a regular smartphone and a standard laptop. Now, a doctor in a village can take an X-ray, snap a photo, and get an instant diagnosis and a "magic heatmap" explanation without needing a single bar of internet signal.

5. The Results: A High-Scoring Student

They tested this system, and the results were impressive:

• Accuracy: It got the diagnosis right about 91% of the time.
• Safety: Most importantly, it rarely missed a case of TB (it caught 98% of them). In medicine, it's better to be a little too cautious and check a healthy person than to miss a sick one.
• Trust: The "magic heatmaps" showed the AI was looking at the lungs, not random parts of the picture.

The Bottom Line

This paper isn't just about a computer program; it's about bringing high-tech healthcare to the places that need it most.

It's like giving a village doctor a super-powered flashlight that can see the invisible signs of disease and explain exactly what it sees, all while running on a battery in a place with no electricity grid. It turns a complex, expensive medical process into something simple, fast, and accessible for everyone.

Technical Summary

1. Problem Statement

Tuberculosis (TB) remains a leading cause of death globally, with a disproportionate burden on low- and middle-income countries (LMICs). While chest radiography is a primary screening tool, its effectiveness is hindered by:

• Resource Limitations: Lack of access to radiological expertise and laboratory testing in rural areas.
• Deployment Barriers: Most existing AI solutions rely on cloud computing or high-performance hardware, making them unsuitable for offline, bandwidth-constrained environments.
• Lack of Explainability: Many deep learning models act as "black boxes," providing classification results without visual evidence, which limits clinician trust and adoption.

The study aims to bridge these gaps by developing a system that is accurate, explainable, and deployable offline on resource-constrained devices (mobile and desktop).

2. Methodology

The study followed an end-to-end AI workflow involving data preparation, model development, explainability integration, and edge deployment.

A. Dataset and Preprocessing

• Data Source: Publicly available chest X-ray datasets containing two classes: Normal and Tuberculosis.
• Preprocessing: Images were resized to $224 \times 224$ pixels and converted to RGB.
• Augmentation: To prevent overfitting and improve generalization, the training set utilized random horizontal flips, rotations, zooms, and brightness/contrast adjustments.
• Splitting: Data was stratified into 70% training, 15% validation, and 15% test sets to ensure class balance and prevent data leakage.
• Class Imbalance: Class weights were calculated and applied to the loss function to penalize the model more heavily for misclassifying the minority class.

B. Model Architecture

• Base Model: DenseNet121 initialized with ImageNet pre-trained weights (Transfer Learning).
• Structure:
  1. Feature Extraction: The pre-trained DenseNet121 backbone extracts features.
  2. Global Average Pooling: Reduces spatial dimensions to a feature vector.
  3. Regularization: A dropout layer prevents overfitting.
  4. Classification Head: A fully connected layer with a Sigmoid activation function outputs the probability of TB ($0 \leq y \leq 1$).
• Training Strategy: A two-stage approach was used:
  1. Freezing: Only the classifier head was trained while the backbone remained frozen.
  2. Fine-tuning: Upper layers of the backbone were unfrozen and trained with a lower learning rate to adapt features to the specific medical domain.
• Loss Function: Weighted Binary Cross-Entropy (WBCE) with the Adam optimizer.

C. Explainability (Grad-CAM)

• Technique: Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to visualize the regions of the X-ray influencing the prediction.
• Mechanism: It computes gradients of the class score with respect to the final convolutional feature maps, generating a heatmap that highlights anatomically relevant lung regions (e.g., upper/mid-lung fields).

D. Deployment

• Conversion: The trained Keras model was converted to TensorFlow Lite (TFLite) to enable lightweight, on-device inference.
• Platforms: The system was deployed as:
  1. A Windows Desktop Application.
  2. A Flutter-based Mobile Application.
• Capability: Both platforms support offline inference, allowing operation without internet connectivity.

3. Key Results

The system was evaluated on an independent test set and through external image inference.

• Classification Performance:
  • Accuracy: 0.91
  • Precision: 0.89
  • Recall (Sensitivity): 0.94 (Critical for screening to minimize missed cases).
  • F1 Score: 0.91
  • AUC (Area Under Curve): 0.96 (Reported as ~0.9966 in the ROC curve description), indicating excellent discriminative ability.
• Confusion Matrix Analysis:
  • The model correctly identified 514 Normal and 103 Tuberculosis cases.
  • False Negatives: Only 2 cases (High sensitivity).
  • False Positives: 11 cases (Acceptable trade-off for a screening tool prioritizing sensitivity).
• Explainability Validation:
  • Grad-CAM visualizations confirmed the model focused on clinically relevant lung regions (upper and mid-lung fields) for TB cases.
  • Normal cases showed weak or diffuse activation, indicating the model did not rely on artifacts or background noise.
• Deployment Consistency:
  • Predictions and Grad-CAM heatmaps were consistent across both Windows and Mobile platforms, confirming that the TFLite conversion preserved the model's logic and feature representations.

4. Key Contributions

1. Offline Deployability: Successfully converted a complex deep learning model into a lightweight format (TFLite) for real-time, offline use on mobile and desktop devices, addressing the connectivity challenges in LMICs.
2. Explainable AI (XAI): Integrated Grad-CAM directly into the deployment pipeline, providing clinicians with visual evidence (heatmaps) to support decision-making and build trust.
3. End-to-End Framework: The study moves beyond theoretical model training to a fully functional "TBAI Africa" system that includes data preprocessing, training, explainability, and cross-platform deployment.
4. High Sensitivity Screening: The model achieved a recall of 94%, making it highly suitable for screening applications where missing a positive case is the primary risk.

5. Significance and Impact

• Clinical Relevance: The system provides a practical decision-support tool for healthcare workers in remote settings who may lack radiological expertise. The probability output allows for risk stratification and triage.
• Scalability: By enabling offline operation on standard mobile phones and laptops, the solution is scalable to the most resource-constrained environments.
• Trust and Adoption: The inclusion of Grad-CAM addresses the "black box" critique of AI in medicine, showing where the model is looking, which is essential for clinical validation.
• Future Direction: The authors acknowledge limitations (e.g., binary classification only, lack of multi-center validation) and propose future work involving multi-class classification and prospective clinical trials.

In conclusion, this paper presents a robust, interpretable, and accessible AI solution that effectively bridges the gap between advanced deep learning research and practical, life-saving medical applications in high-burden, low-resource settings.