Lung Nodule Detection Based On Deep Learning

As the main sign of early lung cancer,the pulmonary nodules detection plays an important role in the diagnosis of lung cancer.The rapid development of CT technology has effectively improved the detection rate of pulmonary nodule.However,with the continuous development of imaging technology and the continuous increase of relevant clinical requirements,the work of image diagnostic physicians becomes more and more onerous,it is easy to get tired after reading the CT scanning for a long time,which leads to a certain degree of missed diagnosis and misdiagnosis.Therefore,it is urgent to study computer-aided detection and diagnostic technology to assist doctors.In recent years,with the development of neural network and deep learning technology,it has become a hot direction to expand the deep learning technology to the field of medical processing.This paper mainly studies the application of deep learning technology in the detection of pulmonary nodules.It mainly detects the location and size of pulmonary nodules.Before nodule detection,a method of lung parenchyma segmentation based on threshold segmentation and morphological manipulation was proposed.First,each slice in the CT scan was binarized according to a fixed threshold.Then,the binary images were processed successively by clearing border,2D connected region marking,erosion operation,closure operation and hole filling in the 2D morphology operations.At this time,the approximate lung parenchymal area has been segmented.Finally,a 3D connected region marker was used to extract the left lung and right lung regions,and a relatively smooth and impurity-free lung parenchymal region was obtained.On the one hand,lung parenchymal segmentation can provide a reliable detection area for pulmonary nodules detection and reduce false positives caused by the detection of tissues beyond the lung parenchyma by pulmonary nodules detector.On the other hand,it can reduce the running time of detector and accelerate the detection speed.After lung parenchymal segmentation,a pulmonary nodule detection network based on 3D Faster R-CNN was proposed,which adopted the Dual Path Block and approximate U-Net codec structure to make full use of the spatial relationship between all slices of the patient’s single CT scan.First,all slices are input and they were processed into a 3D data through linear transformation and sampling operations.Then,the detection network was trained to learn the 3D characteristics of nodules.Finally,the possibility of nodules in the 3D data block was predicted and the center position and size of the suspicious nodules were obtained by regression method.Dual path connection has the advantages of both residual connection and dense connection,and can fully learn new features on the basis of feature reuse.The nodule detection network designed by this structure has the advantages of fewer parameters and better effect.To verify the effectiveness of the algorithm used in this paper,we validate the algorithm on the LUNA16 public dataset,the highest FROC obtained on the 10 subsets was 0.910,and the average FROC was 0.864,indicating that this algorithm has a good distinction between true and false nodules;the highest recall rate was 0.991,and the average recall rate was 0.951,indicating that this algorithm had a high nodule detection rate.At the same time,the results were compared with the current leading pulmonary nodule detection algorithm.The experimental results show that our approach can achieve the same or even better performance,which proves that the algorithm proposed in this paper has certain practical value and research prospect in the field of pulmonary nodule detection.The whole work in this paper has played a positive role in promoting the detection rate and accuracy of pulmonary nodule lesions by using computer aided detection technology.