Object Detection And Segmentation In Medical Images

With the development of medical imaging technology,medical images have become an important tool for doctors to diagnose diseases.Therefore,the automatic processing and analysis of images is of great significance.Computer aided image processing can reduce the doctors’ repeat labor and increase their efficiency.Furthermore,with the accurate target location and segmentation,diseases diagnosed by computers will be possible and could help hospitals supplement their staffs.Traditional medical image processing algorithms rely heavily on the design of artificial characteristics.While the design of such characteristics is complex and the generalization ability of those model is relatively poor,which may lead it impossible to adapt to the actual application due to the different conditions of data collection and Individual differences.In recent years,deep learning has been making a breakthrough in many fields such as computer vision and natural language processing.More and more researchers have applied deep learning into various tasks of medical images like object detection,target segmentation and disease diagnosis.In this paper,a deep learning algorithm is used to realize the task of automatic detection and segmentation of targets in medical images.The main work and innovation points of this paper are as follows:Firstly,aiming at the task of optic disc location in Fundus image,an end-to-end convolution neural network model is used.The final positioning of the optic disc is realized by training the model firstly to obtain the probability graph which including the position of the disc.Then threshold processing used to reduce interference and then the location of optic disc is finally calculated as the barycenter of probability graph.Compared with other algorithms,the network designed in this paper can locate the optic disc both quickly and accurately for various types of retinal fundus images.The results show that this method has good generalization performance and has better effect than similar algorithm.Secondly,for the task of vascular segmentation of fundus image,an improved vascular segmentation algorithm was proposed in this paper.For the different types of blood vessels in the fundus image,a multi-scale network structure is designed to extract features of both main blood vessels and vessel branches at the same time.Therefore,the segmentation model proposed can achieve good results on all kinds of blood vessels even if they are lower contrast and have few obvious characteristics.In the test stage,the performance of the model is envaluated through multiple evaluation indexes that are widely used in the field of medical image segmentation.The results show that the method proposed by us has a great improvement over the segmentation algorithm of same task.Furthermore,the results generated by our model can achieve comparable effect as the segmentation of human doctor.In addition,we designed a cascade CNN model to segment MRI’s brain glioma accurately and efficiently.According to the different types of tumor,a corresponding segmentation model was trained in different dimensions to improve the comprehensive segmentation effect.Aiming at the problems of large image size and unbalanced data in 3D Medical Image Segmentation task,this paper synthesizes the advantages of 3D and 2D networks and the segmentation results achieve the state of art by cascading method in the case of ensuring the complexity of model calculation.In addition,we have further improved the effect of the model through related post-processing operations such as morphological processing and DENSECRF etc.