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

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) present as ground-glass nodules (GGNs) on CT scans.
• The low contrast, blurred boundaries, and variable appearance of GGNs challenge automated detection systems.

Purpose of the Study:

• To develop a lightweight yet accurate deep learning framework for detecting AIS and MIA.
• To improve the detection of challenging ground-glass nodules (GGNs) in pulmonary computed tomography (CT) scans.

Main Methods:

• Proposed FAD-YOLO (Feature-refinement, Alignment, and Dynamic-sampling YOLO) framework, building upon YOLO12n.
• Introduced three key modules: A2C2f-FRFN for feature refinement, DySample for dynamic upsampling, and TADDH for task-aligned detection.

Main Results:

• FAD-YOLO achieved high precision (93.8%) and recall (93.4%) on an internal test set.
• The model demonstrated comparable or superior performance to existing detectors (YOLOv5n, YOLOv8n, YOLO11n, RT-DETR) with significantly fewer parameters.
• Achieved strong results (mAP@50 = 91.7%) on an external dataset without fine-tuning.

Conclusions:

• FAD-YOLO offers a balanced solution for accuracy, lightweight design, and cross-dataset generalization.
• The model shows potential as an assistive tool for radiologists in AIS/MIA discrimination, especially on resource-constrained devices.