Research On Intelligent Diagnosis Assisted Of Pulmonary Nodules Based On Deep Learning

Early detection and diagnosis of pulmonary nodules is essential to improve the survival rate of lung cancer patients.With the wide application of artificial intelligence technology in the medical field,the pulmonary nodule auxiliary diagnosis system based on deep learning can help radiologists better complete the diagnosis of lung CT images,significantly reduce the burden of radiologists,and improve diagnosis efficiency.This thesis aims at the difficulties caused by the different diameters of lung nodules in the detection and classification of lung CT images based on deep learning and the insufficient benchmark datasets.A pulmonary nodule detection and classification algorithm based on2 D convolutional neural network is proposed.And the auxiliary diagnosis system for pulmonary nodules is built to visualize the results of nodule detection and classification,which enhances the practicability of detection and classification algorithms.The main contents are as follows:1.For the detection of pulmonary nodules of different sizes in CT images,especially the detection of small nodules,a lung nodule detection network based on multi-scale feature fusion,multi-attention mechanism,and multi-task is proposed.The proposed model introduces spatial attention and channel attention mechanisms in the feature extraction network,and adds spatial pyramid convolution with self-attention mechanism to construct a multi-task branch with detection and segmentation for lung nodule detection at different scales.On the LUNA16 dataset,experimental results show that the proposed method can achieve a sensitivity of 94.0% at 0.125 false positives per scan and a CPM(Competition Performance Metric)score of 94.9,outperforming many of the current mainstream lung nodule detection models.In addition,the detection sensitivity of this detection model can still reach more than 90% for small nodules.2.To address the current problem of insufficient samples in the classification of benign and malignant lung nodules,an algorithm for the diagnosis and classification of benign and malignant pulmonary nodules based on self-supervised transfer learning is investigated,which extracts multi-scale features through a multi-convolution process and captures more high-level features and semantic information based on residual blocks and a self-attention mechanism.On the LIDC-IDRI dataset,our model can achieve 94.10% classification accuracy of pulmonary nodules with an AUC of 0.98.The experimental results demonstrate the superiority and effectiveness of the proposed model in this thesis at a lower computational cost.3.An auxiliary diagnosis system for pulmonary nodules is designed,which fully visualizes the detection of pulmonary nodules and the results of benign and malignant diagnosis and assists radiologists in their diagnosis decisions.