Research On Aided Diagnosis Of Pulmonary Nodules Based On Deep Learning

Lung cancer is the most deadly and common lung tumor.Its early manifestation is the proliferation of pulmonary nodules.Chest X-rays can identify a variety of lung diseases.However,there is an imbalance between the number of professional radiologists and the large number of chest X-ray images produced each year,resulting in an imbalance in the doctor-patient ratio.Aiming at the subject of auxiliary diagnosis of pulmonary nodules,a computer-aided diagnosis system for lung region segmentation and lung nodule detection based on chest X-ray images was developed by using deep learning technology.The main contributions of this paper are as follows:In-depth research on lung region segmentation based on CXR images is carried out to solve the following difficulties and challenges: 1)inaccurate edge segmentation,2)poor segmentation results due to lesions,and 3)lack of ability to utilize multi-scale effective information.In this paper,a new mask attention network,referred to as MA-Net,is designed.An edge-assisted computing module is designed to guide the edge prediction of the lung region,and the mask attention module is used to enhance the semantic features of the lung region to reduce the interference of the diseased region.The lung region prediction results at each scale are used as optimization targets using a multi-scale aggregate loss function.In this paper,MA-Net is compared with baseline methods and advanced lung region segmentation methods on three datasets,JSRT,Montgomery and Shenzhen.The experimental results show that MA-Net has advantages in ACC,Dice and Jaccard.The ACC,Dice and Jaccard of MA-Net using Resnet50 as the encoder module reach 98.95%,98.11%and 96.14% respectively.In-depth research is carried out on the precise localization of pulmonary nodules based on CXR images,and the following difficulties and challenges are solved: 1)the feature extraction ability is insufficient,and 2)the targets of pulmonary nodules are small and vary in size.This paper proposes a new dual-path fusion attention network,referred to as DFA-Net,and designs an encoder-decoder structure to achieve lung nodule detection.A dual-path feature extraction module is proposed to fuse the advantages of skip connections and dense connections to improve the performance of the network.Fusion of low-level and high-level feature maps using a fusion attention module further enhances the effective features of lung nodules,and lung nodules of different sizes are detected using three anchors of different sizes in the multi-scale prediction module.In this paper,DFA-Net is compared with advanced lung nodule detection methods on the JSRT dataset.The experimental results show that DFA-Net has advantages in both FROC and CPM.The d setting of the dual-path feature extraction module It has the best detection performance when it is 0.5,its CPM score is 0.826,and its sensitivity reaches 97.5% in the case of 5 false positives per image.In addition,ablation experiments are performed on the fused attention module,and the experimental results show that the CPM score of DFA-Net is improved by 4.3% using the fused attention module.