Research On Lung Nodule Detection Algorithm Based On Hybrid Model Of CNN And Transformer

Lung cancer is a common malignant tumor worldwide,causing high morbidity and mortality rates.In clinical practice,lung nodules are early manifestations of lung cancer,and doctors can generate CT images of the chest to detect,evaluate,and diagnose them.However,identifying lung nodules may be challenging as they may adhere to other tissues,leading to fatigue among doctors and potentially resulting in missed or incorrect diagnoses.To address these issues and to enable accurate and automatic detection of lung nodules,this paper proposes a lung nodule detection algorithm based on a combination of CNN and Transformer.Specifically,the Swin Transformer which is fully based on Transformer is used to extract features,while the YOLOv5s network architecture handles feature fusion and object detection.The algorithm achieves promising results and was programmed for easy-use by physicians without programming experience.The main research contributions are as follows:(1)A collaborative hospital dataset was developed using real case data from 2011 to 2020.Three professional radiologists annotated the images,and different types of lung nodules were resampled to address the issue of sample imbalance.Consequently,a dataset containing 17,178 JPG format images categorized as 5,038 adenocarcinoma images(A),4,813 small cell carcinoma images(B),3,311 large cell carcinoma images(E),and 4,016 squamous cell carcinoma images(G)was established.(2)Two algorithms,namely YOLOv5s and DETR,were utilized for lung nodule detection.Both algorithms were validated through comparative analyses and manually labeled results,showing their feasibility and effectiveness when detecting various types of lung nodules.(3)The Swin Transformer was introduced as the backbone network for feature extraction,which boosted the learning ability of the YOLOv5s model and enhanced detection accuracy.Comparative and ablation experiments were conducted with other traditional Convolutional Neural Network models.(4)A software was developed for use by physicians without programming experience.The algorithm was deployed on Jetson Xavier NX edge computing devices to reduce hardware costs with Pyqt5 utilized for developing the visual interface.