Research On Automatic Diagnosis Algorithm Of Pulmonary Nodules Based On CT Images

It is clear from studies that the incidence and mortality rate of lung cancer is increasing year by year,and it is always at the top of malignant tumors in China.Lung cancer is difficult to be detected in the early stage,resulting in missed treatment time.Currently,several imaging techniques have been applied to the early diagnosis of lung cancer,among which computed tomography(CT)is the most common and efficient imaging-based lung cancer diagnosis method.It has become a popular trend to use deep learning-based computer-aided diagnosis(CAD)systems to help physicians improve the detection rate of pulmonary nodules and the diagnosis rate of lung cancer.The paper proposes an automatic diagnosis method of pulmonary nodules based on CT images to address the problems of low sensitivity,high false alarm rate,and difficulty in diagnosing small nodules in existing CAD systems.Candidate nodules are first extracted by a nodule detector based on a 3D multi-scale dual-path network,and then the detected candidate nodules are cropped and input to a separate classification network for benign and malignant diagnosis.The research content of this paper is mainly divided into the following three parts.This paper proposes to use a threshold-based lung parenchyma segmentation method combined with morphological operations to achieve lung parenchyma segmentation,where the morphological operations include two-dimensional and three-dimensional morphological operations.In this paper,a relatively broad threshold range is selected to obtain a rough lung parenchyma area,which is further processed by morphological operations.The experimental results show that this method can obtain smooth and impurity-free lung parenchyma.Aiming at the characteristics of different shapes and sizes of pulmonary nodules,this paper proposes a pulmonary nodule detector based on a 3D multi-scale dual-path network.First,the dual-path network realizes feature reuse while mining new features,and has good feature extraction ability.Second,considering the ability of the Inception model and U-Net structure to extract multi-scale features,two 3D Inception modules are used as preprocessing blocks,and a U-shaped network structure is constructed.Finally,considering the imbalance of positive and negative samples in the lung nodule detection task,a new loss function is introduced to make the network more inclined to learn more difficult samples.The detection task uses FROC and CMP as the evaluation indicators.The results based on the LUNA16 dataset show that the detector ensures a high sensitivity rate while maintaining a high accuracy rate,and can detect lung nodules as much as possible.Since pulmonary nodule diagnosis requires more refined features,it can be achieved by focusing on the nodule.This paper proposes a lung nodule classifier based on the improved DPN module.Inspired by the attention mechanism in enhancing the feature extraction ability,this paper introduces the SE module based on DPN and changes the residual connection mode in the DPN module to the second-order response fusion,which increases the nonlinear fitting ability of the network.At the same time,the propagation of responses across branches is facilitated.Due to the good classification performance of GBM,the depth feature extracted by 3D convolutional neural network and effective artificial features such as nodule diameter and nodule pixel are used as the input features of the GBM classifier,and good performance is achieved.