RadTriage: Vision-Language Zero-Shot Radiology Triage
A CLIP-style vision-language model fine-tuned on paired chest X-ray/report data for zero-shot urgency triage of incoming studies.
Overview
RadTriage builds on the CLIP reproduction to align chest X-ray images with radiology report text, enabling zero-shot ranking of incoming studies by likely urgency using natural-language prompts (e.g., "an urgent finding" vs "a normal study") rather than a fixed label set.
Architecture
Image encoder initialized from the earlier ResNet-50 reproduction, text encoder fine-tuned on radiology report language; contrastive fine-tuning on paired image-report data followed by prompt-based zero-shot scoring at inference.
Tech Stack
Screenshots
Lessons Learned
Domain-specific vocabulary in radiology reports (abbreviations, negation patterns like "no acute findings") required text-encoder fine-tuning on in-domain text before zero-shot prompting became reliable — the general-purpose pretrained encoder alone under-performed noticeably.
ls -la ./related-projects/
Edge-NIDS: On-Device Intrusion Detection for IoT
A quantized CNN intrusion-detection model deployed to an ARM Cortex-M microcontroller for real-time IoT traffic monitoring without cloud connectivity.
FedDiagnose: Cross-Hospital Federated Diagnostic Modeling
A simulated multi-hospital federated learning platform for chest X-ray classification, keeping patient imaging data local to each institution.
ReproBench: Automated Paper Reproduction Scorecard
A tooling suite that runs reproduced model checkpoints against held-out benchmarks and auto-generates a comparison scorecard against reported paper results.