Improved Yolov8-Based Algorithm for Pulmonary Nodule Detection

Pulmonary nodule identification serves a crucial function in lung cancer screening, which employs low-dose CT scanning. An improved method named YOLOv8 + SE + Soft -NMS (YOLO8 - xse) based on YOLOv8 is proposed to tackle the limitations in small object detection accuracy and bounding box prediction precision within the YOLO framework. This method aims to detect pulmonary nodules more effectively and achieve sustained convergence of detection bounding boxes. The proposed approach replaces the YOLOv8 backbone with MobileViT (Mobile Vision Transformer), a network known for its powerful feature extraction and representation capabilities. MobileViT enables the direct learning of pulmonary nodule characteristics and detection details from raw images. Additionally, the model integrates the SE (Squeeze-And-Excitation) module into the head network. This channel attention mechanism dynamically evaluates and emphasizes the importance of individual feature channels, improving the identification of critical regions. Considering the small size of pulmonary nodules and their tendency to overlap with surrounding structures, it is Soft-Nms that replaces the traditional non-maximum suppression algorithm. Soft-Nms retains valuable detection information, reducing missed detections and false positives associated with standard NMS. It is through testing on the LUNA16 pulmonary nodule dataset that the YOLOv8-xse model demonstrates significant improvements, with a 3.7% increase in precision, a 3.5% gain in recall, and a 3% rise in mean average precision (mAP:0.5) at a confidence threshold of 0.5. The model also demonstrates a significant reduction in computational load and parameter count, making it highly effective for real-time medical applications.