SilicoPneumoDx

Inspiration

Doctors at Indian mine-site health camps cannot reliably distinguish silicosis from tuberculosis on a chest X-ray — the two diseases produce visually overlapping nodular/fibrotic patterns. A 2025 systematic review found ~41% TB prevalence among silicotic individuals in India, and current AI/CAD tools are TB-only, which actively worsens misdiagnosis in silica-exposed populations. Meanwhile, workers wait 6–8 months for compensation certification even after a positive finding. We wanted to build something that attacks both the diagnostic gap AND the paperwork gap that makes a correct diagnosis meaningless to an informal worker.

What it does

SilicoScan AI is a point-of-care tool for doctors at mine-site health camps. The doctor uploads a chest X-ray plus a short structured patient history, and within seconds gets:

A four-class probability score: Normal / TB / Silicosis / Silico-TB overlap
An ILO-standard profusion grade for silicosis severity
A Grad-CAM heatmap showing which lung zones drove the classification
An auto-drafted compensation/certification packet, pre-filled and routed toward the district Pneumoconiosis Board

How we built it

We designed a multimodal pipeline: a U-Net + ResNet-34 model segments the lung fields, a DenseNet-121 backbone (pretrained via 3-stage transfer learning — ImageNet → NIH ChestX-ray14/CheXpert → TB datasets → silicosis/silico-TB data) extracts features, and a region-based attention layer scores 6 lung zones independently to catch the spatial patterns that distinguish silicosis (upper-zone, symmetric) from TB (apical, asymmetric). A fusion layer combines the image embedding with structured clinical data (occupational history, symptoms, biomarkers) before the final classification head. Explainability (Grad-CAM) and an LLM-based compensation-agent layer for auto-drafting government forms round out the system. Frontend is an offline-first PWA to handle poor connectivity at remote mine sites.

Challenges we ran into

The biggest challenge is labeled data scarcity for silico-TB overlap cases — this is the bottleneck named across every source we researched. We addressed it with staged transfer learning so the model doesn't need thousands of silico-TB images to converge, just hundreds, after building general chest X-ray understanding on large public datasets first. We also had to design around the documented risk that TB-only CAD tools carry population-specific bias when applied to silica-exposed workers.

What we learned

That the technical modeling problem is actually the easier half — the harder problem is the "last mile" of getting a correct diagnosis to actually translate into a worker receiving compensation, which is why we built the compensation-agent layer as a first-class part of the system rather than an afterthought.

What's next for SilicoScan AI

Partnering with district health camps (Jodhpur, Karauli, or Dholpur) to access real silicosis-labeled cohorts, piloting doctor-in-the-loop testing, and integrating the compensation-agent auto-routing with district labour office / ESI systems.