Pulmonary Nodule Detection Based On Deep Learning

Lung cancer is the cancer with the highest incidence rate and mortality,which seriously threatens human life and health.As early lung cancer has no obvious symptoms,it is easily ignored by patients,and when the disease deteriorates to the middle and late stage,the best time for treatment has been missed.Early cancer detection is the most direct way to improve patient survival,so early screening for lung cancer is becoming increasingly important.Pulmonary nodules are early symptoms of lung cancer,and the detection of Pulmonary nodules is a critical step in diagnosing lung cancer.Therefore,using lung computed tomography images to assist radiologists in the rapid and accurate detection of lung nodules through computer-aided diagnostic systems is essential.With the rise of deep learning in recent years,convolutional neural networks have been widely used in pulmonary nodule detection tasks.However,the accurate detection of pulmonary nodules is a highly challenging task due to their small size and variable morphology.This thesis uses a three-dimensional convolutional neural network as the basic structure,combined with methods such as one-shot aggregation,attention mechanism,and multi-scale feature fusion,to propose a pulmonary nodule detection network model,as follows:(1)A pulmonary nodule detection network model based on a one-shot aggregation and attention mechanism is proposed.Firstly,according to the needs of the pulmonary nodule detection task,the U-shaped encoder-decoder structure is used as the feature extraction network,and the 3D region proposal network is used as the detection network.Then,according to the characteristics of pulmonary nodules,a one-shot aggregation module was proposed as the fundamental component of the feature extraction network so that the network could efficiently aggregate the features of the middle layer,retain more diversified information,and avoid the loss of nodule features as much as possible to improve the detection accuracy of nodules.Then,based on the one-shot module,an attention mechanism is added to filter redundant information so that the network pays more attention to the characteristic of pulmonary nodules and suppresses irrelevant information.Finally,the loss function used and the specific network structure are determined through experimental comparison,and the pulmonary nodule detection models based on two-dimensional and three-dimensional convolutional neural networks are compared.The results show that the proposed pulmonary nodule detection model achieves the highest CPM(Competition Performance Metric)value of 0.860.(2)A multi-scale feature aggregation module is proposed to improve the detection effect.Firstly,the multi-scale problem of pulmonary nodules was analyzed.Since the detection network of pulmonary nodules only uses a single scale for prediction and cannot effectively detect nodules of all scales,a multi-scale feature fusion module was added to the detection network of pulmonary nodules to integrate features of different levels.Secondly,the information attributes contained in the features of different levels differ.The high-level features contain more complex semantic information,while the low-level features contain specific location information,and their importance differs.Therefore,features at different levels are fused through the attention mechanism to select important features.Finally,considering that the surrounding tissue information is also conducive to detecting lesions,a receptive field block is added after feature fusion to increase the receptive field to improve the detection effect of the network further.The experimental results show that the pulmonary nodule detection network with a multiscale feature aggregation module performs better than other algorithms,achieves the highest sensitivity at 0.125 FP per scan,and has a higher CPM value,up to 0.903.