LungNoduleAgent
A four-stage multi-agent pipeline, presented at AAAI 2026, that classifies lung adenocarcinoma nodules from CT images and reports through nodule detection, sizing, report generation and a multi-agent clinical debate.
LungNoduleAgent classifies lung adenocarcinoma nodules into AIS, MIA or IAC categories by mirroring how a radiology team reaches a diagnosis: a nodule-detection stage, a sizing stage, report generation, and a multi-agent clinical debate stage that discusses the case before a final classification. Each stage can be independently enabled or disabled in a single configuration file.
The released detection component, Nodule Spotter, is a backend-independent mixture-of-experts pipeline combining three lung-nodule detection models, RetinaNet, Faster R-CNN and DINO, whose configs and checkpoints are published on Hugging Face with pinned artifact revisions and verified checkpoint hashes. The project can run from existing masks and reports without a GPU, from CT images alone with a GPU for detection, or in batch over a JSONL file of cases, and reads model API keys from environment variables.