Detection Of Pulmonary Nodules In Computer Tomography Based On Neural Network

Lung cancer is one of the most deadly cancers over the world.Early diagnosis of lung cancer is crucial to the treatment and the increase of survival rate of patients.Manual diagnosis depends on doctors’ experience,and there are some problems such as long time-consuming and low detection accuracy.Therefore,computer-aided diagnosis technology is applied to medical images.Pulmonary nodule detection models based on neural network are mainly categorized into three-dimensional（3D）model and two-dimensional（2D）model.The former may not have enough images to accurately construct three-dimensional nodules;the latter usually processes chest images slice by slice,and its detection method is more in line with doctors’ reading habits,but has the following problems leading to low detection accuracy:（1）Imaging outside the lungs interferes with nodule detection;（2）Small nodules are difficult to detect because of their low brightness and poor features;（3）Some nodules are attached to other organs in the lungs,which are easily confused with the background during detection.In this thesis,an automatic detection technology for pulmonary nodules is investigated to address the above problems.For problem（1）,a computed tomography（CT）image pre-processing method is designed based on the morphological image processing methods and the basic principle of threshold segmentation in this thesis.Simulation results show that the method improves the sensitivity by 1-3% for each false positive condition.For problem（2）,two shallow convolutional layers are added to the SSD（Single Shot Detector）base network as output layers in this thesis.The optimization method of multi-layer feature map combination is designed to determine the effect of different scale feature map on pulmonary nodule detection.Meanwhile,a variety of improved schemes of default box scale factor,number,and aspect ratio are designed in this thesis to make the size of the default boxes generated in each layer of the feature map more match the size of the pulmonary nodule while ensuring that the receptive field is sufficient.Simulation results show that the sensitivity of the improved model could be increased by 3% under high false positive rate.For problem（3）,the attention mechanism is adopted in this thesis.Two optimization schemes are designed in this thesis: first,an attention module is added after all feature output layers to improve the overall feature extraction ability of the model and to filter all features;second,for small targets in pulmonary nodule detection,an attention module is inserted after the shallow convolution layer of the model to optimize the feature selection for small targets.The simulation shows that the competition performance metric（CPM）value of the model can exceed 0.9 by reasonably selecting the optimization scheme and attention module.Combining the above improved methods,the final neural network model can reach a sensitivity of 96% and a CPM value of 0.922 on the LUNA16 dataset and a CPM value of 0.6 on the PN9 dataset.The results show that the model excels is outstanding in2 D models and has detection performance comparable to mainstream 3D models under the condition of lower complexity.