Research on Optimization of YOLOv8 Pulmonary Nodule Detection Algorithm Based on Deep Learning

• Published in: 2024 5th International Conference on Electronic Communication and Artificial Intelligence (ICECAI) pp. 782 - 788
• Main Authors: Ma, Xingxing, Jia, Jianxin, Yan, Xusheng, Wang, Xinxin, Song, Yang, Zhu, Guiyao, Gao, Yunfei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 31.05.2024
• Subjects: Accuracy; Computational modeling; Computed tomography; Lung; Lung cancer; Neural Network; Pulmonary Nodule; Semantics; Target Recognition; Training; YOLOv8
• DOI: 10.1109/ICECAI62591.2024.10675286

Accurate classification of benign and malignant pulmonary nodules in computed tomography (CT) is of great significance for the early diagnosis of lung cancer. Nevertheless, a host of issues, such as the complexity of the background of pulmonary nodules in CT images as well as the incomplete extraction of image features, have brought some troubles to the detection and classification of benign and malignant pulmonary nodules. Against this backdrop, given the low precision and accuracy of the traditional model in the identification and detection of pulmonary nodules, this research proposes a YOLOv8-Small detection model based on YOLOv8 algorithm optimization. More specifically, first and foremost, the GAM attention mechanism module is added in the Neck section to enhance the semantic and location information in the features, thereby improving the feature fusion ability of the proposed model. Secondly, to avoid the loss of semantic information induced by different scales for detecting small targets, a small target detection layer is added to enhance the fusion of deep semantic information and shallow semantic information. Lastly, the GIoU boundary loss function is employed to replace the original loss function, thus improving the bounding box regression performance of the network. The experimental results demonstrate that the recall and the mean average precision (MAP) exhibited by the improved model are both improved by 3% compared with those in the original model. Simply put, this model is significantly superior to the contrast algorithm in improving the detection accuracy while fulfilling the requirements of edge computing equipment. Hence, it is of practical application value.