The Research Of A Pulmonary Nodule Detection Method Based On The Conventional Neural Network

Lung cancer is one of deadly cancers.Its inchoate symptom is pulmonary nodule.By receiving definitive therapy in the period of pulmonary nodule,patients' survival rate can be improved.With the development of technologies,the detections of pulmonary nodule can be achieved in many ways.This paper presented a novel pulmonary nodule detection method based on the conventional neural network(CNN).The detecting method introduced in this paper can be divided into several components:(?)image preprocessing.This step includes transforming an image's pixel value into Hounsfield unit.(?)Lung segmentation.In this component,two segmentation methods are used:one is general morphological operation and the other is watershed algorithm,(?)Unet model(based on CNN)training.Optimization of the model parameters by inputting the segmented images to the Unet models.Through a validation,I evaluated two conventional operators(3*3 and 5*5)of Unet model and selected an appropriate detection model.This appropriate model was then applied to detect pulmonary nodule and to produce two IOU values which refered to different segmentation methods.Then I conducted a performance comparison of the two methods.For Unet model,the conventional operator of 3*3 has a higher IOUvalue than the one of 5*5.This may be caused by the small size of medical image features that contain clinical significance.A conventional operator with large size is prone to lose information and deteriorates the detection precision.According to the results of detecting pulmonary nodules,the watershed algorithm has a better performance,because this method uses the BlackHat algorithm to further process the pulmonary contour,rather than simple operations,e.g.erosion and dilation.The watershed algorithm is a time-consuming and complex process,and requires a certain level of computing power,although it can improve the success rates.General morphological segmentation operations have the merits of simple processing and fast detection.Hence,doctors are supposed to select the segmentation method according to the practical conditions.