Introduction
Pneumoconiosis, a group of irreversible occupational lung diseases including silicosis and "black lung," remains a significant public health crisis, particularly in the mining and construction sectors of developing nations like India. These conditions are often misdiagnosed as Tuberculosis (TB) due to their visual similarity on standard chest X-rays (CXRs). This diagnostic challenge is compounded by a severe shortage of expert radiologists in remote mining regions. While AI has advanced significantly in detecting common respiratory diseases, occupational lung diseases are frequently under-represented in standard datasets, leading to suboptimal performance by general-purpose medical AI models.
The Proposed AI-Assisted Pipeline
To address these gaps, researchers have developed a multi-stage "Digital Diagnostic Assistant" designed to provide explainable and actionable insights for healthcare workers. The pipeline moves beyond simple binary classification to offer a comprehensive diagnostic tool.
The framework consists of several key components:
- Preprocessing & Segmentation: Enhances image quality and isolates lung regions to reduce noise.
- Abnormality Localization: Utilizes Mask R-CNN to identify and bound specific pathological findings such as nodules or opacities. This "local labeling" ensures clinicians can verify the AI's focus areas.
- Disease Prediction: Employs CheXNet to categorize images into four distinct classes: Normal, TB, Silicosis, or Silicotuberculosis.
- Automated Reporting: A GPT-based Transformer architecture converts localized findings into natural language medical reports, bridging the expertise gap for non-specialist healthcare workers.
The Sirohi Dataset
A critical contribution of this work is the introduction of the Sirohi Dataset, specifically tailored to the Indian context. It comprises 4,077 CXR samples collected from primary healthcare centers in the Sirohi district of Rajasthan. The dataset was meticulously annotated by expert radiologists from Tele-Radiology Solutions Bangalore and AIIMS Jodhpur. It includes both global disease labels and localized annotations for findings like lung opacities and pleural effusion, providing a robust foundation for training specialized models.
Experimental Results and Insights
Baseline evaluations were conducted using three prominent deep learning architectures: ResNet18, DenseNet121, and InceptionV3. The results highlight the inherent difficulty in distinguishing between pure Silicosis and pure TB, a phenomenon referred to as the "TB Mimicry Trap."
As shown in the performance table, while models achieved high AUC scores for "Normal" cases (up to 0.91), performance dipped significantly for TB (as low as 0.71). This underscores the necessity of the specialized, explainable architecture proposed in this study. The high inter-class similarity between early-stage silicosis and TB remains the primary hurdle for automated systems.
Conclusion and Future Directions
While the proposed framework shows promise, several challenges remain. Deep learning models require even larger, more diverse datasets to achieve clinical-grade reliability. Furthermore, deploying AI in rural settings necessitates overcoming infrastructure constraints such as inconsistent power and internet connectivity.
Future work will focus on refining the model's ability to distinguish between various pneumoconiosis variants and integrating the tool into broader telemedicine frameworks. By combining automated localization with natural language reporting, this AI-assisted approach aims to empower local healthcare providers and improve outcomes for vulnerable worker populations.
Summary updated