3D multi‐resolution deep learning model for diagnosis of multiple pathological types on pulmonary nodules

• Published in: International journal of imaging systems and technology Vol. 32; no. 1; pp. 74 - 87
• Main Authors: Fu, Yu, Xue, Peng, Zhao, Peng, Li, Ning, Xu, Zhuodong, Ji, Huizhong, Zhang, Zhili, Cui, Wentao, Dong, Enqing
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.01.2022; Wiley Subscription Services, Inc
• Subjects: Accuracy; Computed tomography; Datasets; Deep learning; Image classification; Image databases; Lungs; Machine learning; Medical imaging; multiple pathological types; multi‐resolution method; Nodules; pulmonary nodules
• ISSN: 0899-9457; 1098-1098
• DOI: 10.1002/ima.22642

To accurately diagnose multiple pathological types of pulmonary nodules based on lung computed tomography (CT) images, a multi‐resolution three‐dimensional (3D) multi‐classification deep learning model (Mr‐Mc) was proposed. The Mr‐Mc model was constructed by using our own constructed lung CT image dataset of pulmonary nodules with clinical pathological information (LCID‐CPI), which can accurately diagnose inflammation, squamous cell carcinoma, adenocarcinoma, and other benign diseases. In order to process nodules with different sizes, a multi‐resolution extraction method was proposed to extract 3D volume data with different resolutions from lung CT images. The Mr‐Mc was composed of three different resolution networks, each of which has input volume data of a specific resolution. Experiments showed that the constructed Mr‐Mc model can achieve an average accuracy of 0.81 on LCID‐CPI. Besides, the Mr‐Mc model can also achieve a high accuracy of 0.87 on the Lung Image Database Consortium and Image Database Resource Initiative dataset.