Abstract
Automated diagnosis of tuberculosis (TB) from chest X-Rays (CXR) has been tackled with either hand-crafted algorithms or machine learning approaches such as support vector machines (SVMs) and convolutional neural networks (CNNs). Most deep neural network applied to the task of tuberculosis diagnosis have been adapted from natural image classification. These models have a large number of parameters as well as high hardware requirements, which makes them prone to overfitting and harder to deploy in mobile settings. We propose a simple convolutional neural network optimized for the problem which is faster and more efficient than previous models but preserves their accuracy. Moreover, the visualization capabilities of CNNs have not been fully investigated. We test saliency maps and grad-CAMs as tuberculosis visualization methods, and discuss them from a radiological perspective.
Introduction
Tuberculosis remains a critical global health threat, particularly in low-resource regions where access to expert radiologists is limited. While deep learning offers a solution for automated screening, current state-of-the-art models are often "borrowed" from natural image classification tasks (e.g., ResNet or AlexNet). These architectures are frequently oversized for medical imaging, leading to high hardware costs and a tendency to overfit on smaller clinical datasets. This research introduces a lightweight CNN architecture specifically designed to be efficient enough for deployment on low-cost hardware like a Raspberry Pi without sacrificing diagnostic accuracy.
Methodology
The proposed architecture is a custom CNN consisting of five convolutional blocks designed for parameter efficiency. Each block utilizes two $3 \times 3$ convolutions followed by a max-pooling layer. To ensure stable training and better gradient flow, the researchers incorporated shortcut connections via $1 \times 1$ convolutions, a technique inspired by ResNet but scaled down significantly.
Key design choices include:
- Global Average Pooling (GAP): Replacing traditional, memory-intensive fully connected layers with GAP reduces the parameter count and helps prevent overfitting.
- Data Augmentation: Elastic deformations were applied to training data to simulate natural anatomical variations, forcing the model to learn robust, generalizable features.
- Interpretability Tools: The study employed Saliency Maps and Grad-CAM to visualize the decision-making process, allowing for radiological verification of identified lesions.
Results and Performance
The model was evaluated using three primary public datasets: Montgomery County (MC), Shenzhen (SZ), and a Belarus dataset, as well as a combined (CB) dataset to test robustness. Despite having only 230,000 parameters—a massive reduction compared to the 60 million in AlexNet—the model achieved highly competitive results.
On the combined dataset, the model reached an AUC of 0.925 and an accuracy of 86.2%. Computationally, it requires only ~350 MFLOPs, representing a 4-fold to 30-fold reduction in operations compared to standard architectures. In practical terms, inference takes just 5–6 milliseconds on entry-level GPUs, making it ideal for field deployment in mobile clinics.
Discussion and Interpretability
A significant finding of the study was the effectiveness of different visualization techniques. While Grad-CAM is popular in general computer vision, the authors found that Saliency Maps provided better resolution for identifying the fine-grained, patchy opacities characteristic of TB. Visualization also revealed that the network relies heavily on the edges between the mediastinum and lung tissue for its predictions.
However, some limitations remain. The model occasionally produces false positives due to "bone superposition," where overlapping ribs or clavicles create shadows that mimic TB symptoms. Future work will focus on incorporating pre-training and generating automated textual descriptions of symptoms to further assist clinical officers.
Conclusion
By optimizing network architecture specifically for chest X-rays, it is possible to create deep learning models that are both highly accurate and exceptionally lean. This "right-sized" approach addresses the dual challenges of hardware constraints and data scarcity, providing a viable path for automated TB screening in the world's most vulnerable regions.
Summary updated