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

Adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) typically present as ground-glass nodules (GGNs) on pulmonary computed tomography (CT), and their low contrast, blurred boundaries, and morphological variability pose substantial challenges to automated detection.

Using YOLO12n as the baseline, we propose FAD-YOLO (Feature-refinement, Alignment, and Dynamic-sampling YOLO), a lightweight yet accurate detection framework with three task-oriented improvements: an A2C2f-FRFN module that replaces the standard feed-forward network within the Area-Attention C2f module with a Feature Refinement Feed-forward Network; a DySample dynamic upsampling module that replaces nearest-neighbor interpolation with content-aware learnable offsets; and a TADDH task-aligned dynamic detection head that combines a shared group-normalized convolution, deformable sampling in the localization branch, and layer-attention reweighting to alleviate classification-localization misalignment.

On an independent internal test set of 317 images, FAD-YOLO achieved a precision of 93.8%, a recall of 93.4%, an mAP@50 of 93.6%, and an mAP@50-95 of 70.8%, while reducing parameters by 18.7% relative to the baseline at the cost of a 12.5% increase in GFLOPs (from 6.4 to 7.2). Under identical settings, it performed comparably to or better than YOLOv5n, YOLOv8n, YOLO11n, and RT-DETR-R18, and outperformed the much larger RT-DETR-R50 on all accuracy metrics despite using approximately one-twentieth of its parameters. On an external Mendeley Data test set, the model achieved mAP@50 = 91.7% and mAP@50-95 = 68.2% without any fine-tuning.

FAD-YOLO achieves a favorable balance among accuracy, lightweight design, and cross-dataset generalization, and may serve as a candidate for further prospective validation as an aid to radiologists in AIS/MIA discrimination on resource-constrained clinical devices.