ABSTRACT Early detection of pulmonary nodules is crucial for timely lung cancer diagnosis. Conventional CNN‐based detectors struggle to capture long‐range dependencies, while pure Transformer models require extensive pretraining. To address these issues, we adopt the YOLO family as the base framework due to its one‐stage, end‐to‐end design, strong multi‐scale and small‐object detection, and computational efficiency suitable for clinical deployment. Building on this foundation, we propose H 2 ‐YOLO, a hybrid detection framework that combines global context modelling with refined local feature extraction. It employs a MobileNetV4 backbone for lightweight yet expressive feature representation. A Contextual Fusion Structure with a pyramid sparse transformer replaces the conventional PANet neck for global context‐guided multi‐scale fusion, and the asymmetric feature refinement module enhances boundary and texture representation. The RQA‐Head integrates shared convolution with localisation quality estimation to improve parameter efficiency and align confidence with localisation. Designed for resource‐constrained clinical environments, H 2 ‐YOLO offers fast inference and heightened sensitivity to low‐contrast nodules. Experiments on LUNA16 and LungCT datasets show that H 2 ‐YOLO achieves better performance than YOLOv11, with increases of 9.5% in precision, 9.1% in recall, and 9.5% in mAP@0.5 on LUNA16, and improvements of 2.5%, 2.1%, and 1.3%, respectively, on LungCT, while maintaining a competitive computational cost.
Li et al. (Thu,) studied this question.