Research On Segmentation Method Of Lung Nodules Based On Deep Learning

Lung cancer is a kind of disease with high incidence,and its lesions are manifested as pulmonary nodules.It is one of the cancers with the highest lethal rate.Finding a lesion at the early stage of cancer for treatment will greatly reduce mortality.Medical CT(Computed Tomography,CT)images are the main modality for the diagnosis of lung cancer.These images provide medical workers with a large amount of diagnostic information to determine the type of cancer.However,manual reading requires medical workers to have strong professional literacy and the diagnosis is timeconsuming,so the research of computer-aided diagnosis(CAD)is of great significance.But in chest CT images,the radiation density of the cancer part is not very different from the background,and the target area is small,and the background area is much larger than the target.These problems will bring certain difficulties to the target segmentation.Based on the above problems,this article mainly did the following research:(1)In view of the problem that the nodules in chest CT images are very close to the radiation density of other tissues,and the characteristics of lung nodules are not obvious,a method of segmenting lung nodules combining Sobel operator and Mask R-CNN(Region-based CNN,R-CNN)is proposed.The method is mainly to use Sobel operator to sharpen the area of high radiation density,and then use the threshold segmentation method to filter out the noise,and then use the enhanced feature map as the input of Mask R-CNN network,and use Res Net50 The network combined with FPN performs feature extraction to avoid the disappearance of features in deep networks.During training,the scale of the anchor box in the RPN(Region Proposal Network,RPN)network is reduced,and the number of anchor boxes generated is reduced to optimize the model.This method has been verified on the LIDC-IDRI dataset,and has achieved a good segmentation effect,and provides very valuable diagnostic information for the doctor’s diagnosis.(2)In the chest CT images of the LIDC-IDRI dataset,the early lung nodules are very small,when training RPN in two-stage detector,most of the anchor boxes are negative samples on the background,and the proportion of positive examples that contain rich nodules’ information is very small.The training ratio between positives and negatives is seriously unbalanced,and the training of a large number of negative examples causes increased noise,which make the model degenerate.The loss function is the core of the neural network model learning,in order to solve this problem,a crossentropy loss function named Focal Loss which along with attention mechanism is proposed to replace the Smooth L1 loss.In order to optimize the model,the Soft NMS(soft non maximum suppression,Soft NMS)algorithm is used to control the positive examples to achieve the purpose of optimizing the model.This method further improves the accuracy of the segmentation at m AP and MIo U metrics after testing on the LIDC-IDRI dataset after testing on the LIDC-IDRI dataset,so this algorithm will provide more reliable information for computer-aided diagnosis.