Research On Automatic Detection Algorithm Of Lymph Nodes Based On Deep Learning

As a lymphatic malignant proliferative disease,lymphoma has a wide range of ages of onset and has different clinical features,which seriously threatens people's health.Screening for lymphoma is mainly carried out by PET/CT detection.A large number of images need to be labeled by a doctor manually,which is cumbersome and time-consuming.The automatic detection of lymphoma can assist the doctor in the examination,and the key technology is the automatic detection of lymph nodes.In this paper,an automatic lymph nodes detection algorithm based on deep learning is proposedThe automatic lymph node detection algorithm proposed includes the lymph node candidate point detection and the lymph node candidate point classification.Firstly,the suspected lymph node candidate targets are extracted from the CT slices,which will then be merged to generate lymph node candidate points.In lymph node classification section,a convolutional neural network is used to classify the candidate points and rejecting the false positive result.The lymph node candidate detection part uses the MaskRCNN target detection network to perform lymph node detection on the patient's CT slice.Lymph node mask is predicted at the same time.The detection results on the slices are combined by a lymph node candidate target merging algorithm.The merging:algorithm combines the candidate boxes belonging to the same target into one candidate point.The lymph node candidate point classification part uses the convolutional neural network to classify lymph node candidate points.The slice of the three axial views of the candidate points is used as the classification network input.The low-level features are extracted separately through multiple input channels and the high-level features are extracted from the merged low-level features by using the residual blocks.Samples are dynamically weighted by the improved Focal loss,which helps training on small data sets.Compared with the previous study,the lymph node candidate point detection algorith,m proposed fewer false positives and reached 78.0%lymph node recall rate at 22FPs/case.The effect of lymph node classification was improved,and a recall rate of 89.3%was achieved at 10FPs/case.