FAD-YOLO: a lightweight feature-refined and task-aligned framework for AIS–MIA discrimination on pulmonary CT

The article presents FAD-YOLO, a new system for detecting adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) on pulmonary CT images. These tumors typically appear as ground-glass nodules and present challenges for automated detection. FAD-YOLO is based on YOLO12n and includes three improvements: a feature refinement module, a dynamic upsampling module, and a task-aligned detection head. On an internal test with 317 images, it achieved 93.8% precision, 93.4% recall, and mAP@50 of 93.6%. The system uses 18.7% fewer parameters than the baseline model and outperforms larger models like RT-DETR-R50 while using approximately one-twentieth of its parameters. On an external test without additional fine-tuning, it achieved mAP@50 of 91.7%, demonstrating good generalization across different datasets.