Auxiliary Diagnosis of Lung Cancer with Magnetic Resonance Imaging Data under Deep Learning

Abstract

This study was aimed at two image segmentation methods of three-dimensional (3D) U-shaped network (U-Net) and multilevel boundary sensing residual U-shaped network (RUNet) and their application values on the auxiliary diagnosis of lung cancer. In this study, on the basis of the 3D U-Net segmentation method, the multilevel boundary sensing RUNet was worked out after optimization. 92 patients with lung cancer were selected, and their clinical data were counted; meanwhile, the lung nodule detection was performed to obtain the segmentation effect under 3D U-Net. The accuracy of 3D U-Net and multilevel boundary sensing RUNet was compared on lung magnetic resonance imaging (MRI) after lung nodule segmentation. Patients with benign lung tumors were taken as controls; the blood immune biochemical indicators progastrin-releasing peptide (pro-CRP), carcinoembryonic antigen (CEA), and neuron-specific enolase (NSE) in patients with malignant lung tumors were analyzed. It was found that the accuracy, sensitivity, and specificity were all greater than 90% under the algorithm-based MRI of benign and malignant tumor patients. Based on the imaging signs for the MRI image of lung nodules, the segmentation effect of the RUNet was clearer than that of the 3D U-Net. In addition, serum levels of pro-CRP, NSE, and CAE in patients with benign lung tumors were 28.9 pg/mL, 12.5 ng/mL, and 10.8 ng/mL, respectively, which were lower than 175.6 pg/mL, 33.6 ng/mL, and 31.9 ng/mL in patients with malignant lung tumors significantly ( $P<0.05$ ). Thus, the RUNet image segmentation method was better than the 3D U-Net. The pro-CRP, CEA, and NSE could be used as diagnostic indicators for malignant lung tumors.