Research On Detection And Segmentation Algorithm For Medical Image Based On Deep Learning

In recent years,the medical data digitalization and big data make more and more intelligent applications on medical images possible,such as the location of tumors and other lesions,the measurement of tissue volume,the study of anatomical structure,etc.As a basic task,detection and segmentation in medical images has an important impact on modern clinical diagnosis and treatment.At present,medical image segmentation is mostly based on traditional digital image processing and machine learning methods,which requires rich experience and complete domain knowledge.In addition,most of the medical images are monotonous in color,unclear in the edge of the lesion and different in shape.It is difficult to describe them by traditional methods such as fixed features.Therefore,in view of the characteristics of medical images and the shortcomings of existing methods,two schemes of convolutional neural networks are proposed,and experimental studies are carried out on the two data sets of nuclear detection and segmentation of microscopic images and liver CT images.In order to solve the problem of the dense distribution of nuclei and small targets in microscopic images,we designed an expand residual network structure for dense small targets detection and segmentation,and improves Mask RCNN model,which can effectively solve the problem of information loss of small targets in deep networks.Experiments on two open data sets of nuclear segmentation show that our model has better recognition and segmentation capability for small and dense targets than other state-of-the-art segmentation algorithms.In order to solve the problem of the complex boundary,blurred edges and inaccurate segmentation of targets in CT images,we designed an Encoder-Decoder structure based on improved UNet,realizing the self-adaptive segmentation algorithm of superficial and deep features,and directly using different scale features to obtain segmentation results.Our model is verified on two data sets of natural images and liver cancer CT images.The experimental results show that our model performs better on PA and other benchmarks than the current popular segmentation algorithms,and has the capability to segment image targets with complex boundary.