A Study On Medical Image Detection Based On Deep Learning Algorithm

Medical institutions produce a large number of medical images every year.These data are in 3D format and often contain a lot of information.At present,these medical imaging data require professional radiologists to spend a lot of time reading and analyzing.The diagnosis result is easy to be influenced by the doctor's own experience and state.In recent years,deep learning has been a great success,especially in the field of computer vision,which provides a new way of thinking and solution for the automatic processing of medical images.However,deep learning is mainly for 2D natural pictures processing,it is difficult to directly expanded into the 3D medical image.Most of the previous testing options for medical imaging required the separation of candidate areas and the screening of false positives in two separate steps,which are not efficient enough both in space and time.This topic puts forward a new medical image detection framework for the above problems,the main work can be summarized as follows:1.A Region Proposal Network(RPN)for 3D medical image processing is proposed.RPN can directly find a large number of suspected lesion boundary frames as candidate areas from 3D medical images.These regional proposals try to contain as many lesion areas as possible,but there may be a large number of false positives.2.In response to the problem of false positives from RPN,the two-stage medical omaging detection network is further proposed.On top of the regional candidate results,the detection network can further optimize the area prediction and disease category of the lesion,and inhibit a large number of false positives from RPN.3.An instance segmentation network for medical imaging is proposed.Compared with the boundary frame prediction of the detection network,the instance segmentation network can directly predict the pixel-by-pixels mask of each lesion area and obtain more accurate lesion information.In order to deal with the regional candidate results efficiently,the author uses CUD A to implement an efficient feature extraction operator based on 3D boundary frame:RoIAlign3D.Unlike RoIPool,RoIAlign3D uses trilinear interpolation to avoid discrete quantization of boundary frames when extracting features,thus avoiding the deviation of position coordinates,which is very important for small target in medical imaging.This paper integrates the above three tasks into a unified framework,called Volume R-CNN.Volume R-CNN can be seen as an extension of the RPN,Faster R-CNN and Mask R-CNN of the 2D R-CNN series,designed specifically for medical imaging data to detect the spatial location of the target at the same time and to obtain an instance segmentation mask.Without whistles and bells,Volume R-CNN has achieved excellent results on the public medical image dataset LUNA16.The authors also analyzed the effectiveness of Volume R-CNN in a detailed controlled experiment.As a general algorithm framework,Volume R-CNN is hopeful to become the basic algorithm of medical image detection.