A Research On Single Point Interaction Based Deep Model For Medical Image Segmentation

With the rapid development of medical imaging technology,various medical images have become important reference information for diagnosis and treatment,and more and more patients benefit from it.However,heavy reading and review work on medical images has become a burden on doctors and reduced their efficiency.Hence,computer-aided medical image processing receives more and more attention.In recent years,thanks to the significant improvement in computing power and deep learning technology,the performance of computer-aided medical image segmentation technology has been continuously improved.Since the medical image data collection and labeling is time-consuming and expensive,most deep learning segmentation models for medical image are patch-based.Patch-based deep models have certain defects:lack of global context information and neglecting the spatial relationship between patches.To overcome the defects mentioned above,two point-based interaction methods are proposed.The details can be concluded as follows:(1)Single point based interaction segmentation method with pseudo mask.In this method,single point interaction and pseudo mask are proposed.To be specific,the central point of prostate is clicked by the physician,which could indicate the approximate location the prostate.Based on the click point and the edge information of other organs,horizontal pseudo mask and vertical pseudo mask are generated.These pseudo masks provide not only the implicit prostate-likelihood maps,but also structure-prior of pelvic.Patches extracted from original image and corresponding pseudo masks are fed into network together.In this way,the network could learn the image feature and global context simultaneously.In addition,to reduce the risk of over-fitting,block dropout mechanism is proposed.The method has demonstrated its effectiveness on the real MR prostate dataset.(2)Single point based interaction segmentation method via sequential learning scheme.In this method,sequential learning segmentation model is proposed.As the method(1),the central point of the object is clicked by the physician.We innovatively model our segmentation task as learning the representation of the bi-directional sequential patches,starting from(or ending in)the given central point of the object.This can be realized by our proposed sequential learning network embedded with a gated ConvRNN unit.In this way,the deep model can take the spatial relationship between adj acent patches into consideration and achieve more accurate segmentation result.The results on real CT kidney tumor dataset and PROMISE12 dataset validate the proposed method could achieve superior performance compared with several related work.