Research On Medical Image Marker Detection Technology Based On Deep Learning

With the continuous progress of medical imaging and computer diagnosis technology,the location of medical image markers is not only a prerequisite for patient diagnosis and treatment planning,but also a prerequisite for many medical image analysis applications including image registration,image segmentation,pathological diagnosis and treatment planning.However,the current practice of medical image analysis is usually manual or semi-manual positioning,which is cumbersome,time-consuming and prone to make mistakes;In addition,due to the differences in human anatomical structure and appearance,it is very likely that similar landmarks will appear in multiple local parts of a structure,making it challenging to accurately and robustly identify medical landmarks.Therefore,the automatic detection of medical image markers has become an essential research direction in the field of medical imaging studies.Combined with the needs of practical medical application scenarios,this paper carried out an in-depth study on the two-dimensional location positioning method of medical marks based on the research method of multi-scale feature fusion,aiming at the common local similarity and positioning accuracy of two medical data sets of cephalic image marks and knee image marks,as well as the respective characteristics of each data set.The main work and contributions are as follows:In order to solve the problem that there are few data sets of Chinese medical landmarks in the location of skull image landmarks,an improved feature fusion network AIW-Net(Attention-Inverted-WNet)based on the attention gate mechanism is proposed.The feature extraction part of the network adopts the lightweight network MobileNetV2,which has been pre-trained by natural images,and significantly reduces the number of parameters while maintaining the same prediction accuracy.The middle module consists of two paths:up-sampling and down-sampling.In the down-sampling path,an improved inverse residual module is used to compensate for the loss of feature information due to the gradual reduction of image resolution.Finally,the intermediate supervision from coarse to fine is introduced in the decoder module to fuse the obtained multi-scale heat map with the feature map,and the attention gate mechanism is used in the jump connection to effectively suppress the response of the background region in the feature map.The test on the ISBI 2015 Grand Challenge benchmark data set shows that the model has high detection accuracy and practicability.For the problem that the hourglass network is easy to lose the bottom details of the image in the training process of knee image marker location,an automatic knee marker detection algorithm based on improved stacked hourglass network and voting regression mechanism is proposed.First,the improved hole space pyramid pooling module(ASPP)is used to complete the feature pre-extraction,and the global information at high resolution is effectively preserved by expanding the receptive field of the network.Then,in the feature extraction and fusion stage,the hourglass network is used to complete the acquisition of multi-level feature information,and the residual structure of coordinate attention is introduced in the channel connection,which effectively captures the local position information and channel information at low resolution,and completes the precise positioning of difficult marker points.Finally,the weighted pixel-level voting regression mechanism is introduced into the prediction module to regress the heat map at the highest resolution and the offset map at the X and Y dimensions.The effectiveness of the model is verified by testing on OA medical key points data set.