Methodology Research Of Nucleus Image Segmentation Based On Deep Learning

Medical imaging-based disease diagnosis plays an increasingly important role in modern medical treatment.It is a time-consuming and labor-intensive task for doctors to observe a large number of medical images for analysis and diagnosis.Computer-aided diagnosis is of great significance for improving the diagnostic efficiency and accuracy of cancer.Nuclear image segmentation is the basis of automatic analysis of medical images and an important part of computer-aided diagnosis.Due to the low contrast of the nuclear image,the large difference in the nuclear distribution and cell adhesion phenomenon,the accurate segmentation of the nuclear image is one of the difficult problems in the automatic analysis of medical images.Based on the deep learning image segmentation algorithm,this paper optimizes it to achieve accurate segmentation of nuclear images.?1?Aiming at the problem of over-segmentation in the segmentation of nuclear images by U-Net,spatial space pyramid pooling module is added in coding network to improve feature extraction capability,and the loss function is improved.Experiments show that the segmentation indicators of improved U-Net network are higher than U-Net and SegNet,significantly higher than the FCN₈s and several traditional segmentation methods.?2?Aiming at the problem that the fully convolutional neural network based on pixel classification can not distinguish different nuclear instances,it is proposed to separate the overlapping nuclei in the post-processing part using the method based on convex defect detection.Experiments show that the method improves the average accuracy and F1 value of the nuclear segmentation evaluation index.?3?Aiming at the necessary of combing post-processing to segment overlapping nuclei for the fully convolutional segmentation network,the instanse segmentation model is constructed based on the Mask R-CNN.Using SE-ResNet as the feature extraction network,the bottom-up feature fusion is added in the feature pyramid network,and the soft-NMS algorithm is used to filter the candidate boxes and target boxes.Experiments show that the model improves the performance of the instance segmentation network to some extent.