Research on YOLOv8 algorithm for pulmonary nodule detection in CT images

Abstract

Pulmonary nodules remain an important radiological marker for early lung cancer and require their accurate detection for early screening and diagnosis. However, because they are small-sized structures that have poor visual contrast and tend to resemble surrounding tissues, modern detection algorithms often find it hard to correctly identify these small lesions. For improved detection accuracy, therefore, this work presents an improved model based on YOLOv8 as its backbone framework. The major improvement is achieved through incorporation of Res2Net as a feature extraction path, thus increasing the network ability to map out multi-scale information and its sensitivity to small nodule structures. Res2Net differs from typical convolutional blocks because it takes a complex approach to grouping and fusion over different scales within each residual block. With Res2Net included as YOLOv8's backbone, its upgraded version further improves recognition accuracy for small-sized targets and ill-defined boundaries while retaining an over-all lightweight structure. The model is trained and validated on CT images processed through standardization and lung parenchyma extraction using the LUNA16 database. Test results show that the upgraded version outperforms the original YOLOv8 in both accuracy and computational efficiency. Additionally, an evaluation is made using alternative prominent detection networks used in practice, including YOLOv5, YOLOv6, and YOLOv3-Tiny. Results show that the upgraded model outcompetes its alternatives in detection efficiency while retaining low computational requirements and shows strong applicability for real-world applications.