SHIELD: A Self-supervised, Silicosis-focused Hierarchical Imaging Framework for Occupational Lung Disease Diagnosis

Yasmeena Akhter, Rishabh Ranjan, Richa Singh, Mayank VatsaIJCAI 2025 · DOI

Introduction

Silicosis is a devastating, irreversible occupational lung disease caused by the chronic inhalation of crystalline silica dust. Prevalent in industries such as mining, construction, and stone-cutting, it represents a significant global health burden. Early detection is critical for patient management, yet traditional diagnostic pathways are often slow, prone to error, and inaccessible in the regions where the disease is most endemic.

The Challenge of Occupational Lung Disease

The diagnosis of Silicosis is currently hindered by three primary obstacles:

  1. Radiological Mimicry: Silicosis often presents with patterns nearly identical to Tuberculosis (TB) on chest X-rays (CXRs), leading to frequent misdiagnosis in high-volume clinics.
  2. Data Scarcity: Deep learning models typically require massive amounts of labeled data, which is notoriously difficult to obtain for niche occupational diseases.
  3. Expert Shortages: There is a severe shortage of specialist radiologists in developing regions to interpret complex scans accurately.
The diagnostic journey from workplace exposure to AI-powered differential diagnosis, highlighting the potential for misdiagnosis between Tuberculosis and Silicosis.

To bridge these gaps, researchers have introduced SHIELD (Self-supervised, Silicosis-focused Hierarchical Imaging Framework), an AI-driven system designed for accurate differential diagnosis in resource-limited settings.

The SHIELD Framework

SHIELD operates through a sophisticated two-phase approach that minimizes the need for large labeled datasets while maximizing diagnostic accuracy.

1. Self-Supervised Pretraining (The "Jigsaw" Task)

Instead of relying on human labels from the outset, SHIELD utilizes Self-Supervised Learning (SSL). The model is presented with segmented lung images broken into a 3x3 grid of shuffled patches. By tasking the model with "solving the puzzle" to restore the original arrangement, it learns the inherent spatial relationships and anatomical symmetry of the lungs—such as the position of the heart relative to the ribs—without requiring a single manual label.

2. Multi-Resolution Hierarchical Strategy

Silicosis pathology manifests at different scales, from fine-grained nodules to large-scale structural scarring. SHIELD addresses this by training models at three distinct resolutions: 1024x1024, 512x512, and 256x256.

  • Pyramidal Fine-Tuning: Models pretrained on the jigsaw task are fine-tuned on a smaller, high-quality labeled dataset.
  • Two-Level Ensemble: The system aggregates predictions from four diverse architectures (ResNet18, VGG16, AlexNet, and DenseNet121) across all resolutions. A learnable weighted score fusion prioritizes the most accurate models for the final diagnosis.

Data and Evaluation

The study utilized a newly curated dataset of 3,044 high-resolution frontal CXRs collected over three years from stone workers in Rajasthan, India. The images were categorized into four groups: Silicosis, Silicotuberculosis (STB), Tuberculosis (TB), and Normal. Ground truth was established by five board-certified radiologists with a high inter-annotator agreement (Cohen’s kappa > 0.85).

Key Insights and Results

SHIELD achieved a diagnostic accuracy of 72%, representing a 20% improvement over traditional transfer learning approaches. The model reached an average AUC of 0.91, performing exceptionally well in identifying "Normal" lungs (0.93 AUC) and Silicotuberculosis (0.92 AUC).

Key findings include:

  • Domain-Specific SSL: Pretraining on lung images significantly outperformed standard ImageNet transfer learning, as the model learned features specific to diseased lung textures rather than general objects.
  • Resolution Synergy: High-resolution models were superior at spotting localized nodules, while lower resolutions helped the model understand global lung symmetry and volume.
  • Interpretability: Using Grad-CAM visualizations, researchers confirmed that the model focuses on bilateral symmetrical patterns in the upper lung zones, mirroring the diagnostic criteria used by human experts.

Conclusion and Future Work

SHIELD provides a scalable, low-cost tool for the early detection of Silicosis, potentially saving lives in industrial settings. While the model still occasionally misinterprets diffuse patterns as TB, future work will focus on validating the framework across more diverse global populations and refining its clinical integration to support frontline healthcare workers.

Summary updated

BibTeX
@inproceedings{akhter2025shield,
  author = {Yasmeena Akhter and Rishabh Ranjan and Richa Singh and Mayank Vatsa},
  title = {{SHIELD: A Self-supervised, Silicosis-focused Hierarchical Imaging Framework for Occupational Lung Disease Diagnosis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1060},
}
Figure