Automatic Detection Method Of Medical Image Anatomical Point

With the development and application of imaging technologies such as imaging omics and computer-aided diagnosis,the location of some key anatomical points in medical images is of great significance for the subsequent modeling and analysis process.At present,in the research of medical image analysis,the detection of anatomical points mostly relies on the manual marking of doctors.This manual detection method not only wastes the precious time and energy of doctors,but also the accuracy of detection results will be affected by some subjective factors such as doctors' working status and clinical experience.Therefore,the research on automatic detection of medical image anatomical points is of great significance to the development and progress of modern medicine.In this paper,an automatic detection method of anatomical points in 3D medical images is proposed based on four anatomical points in head MRI which can be used for image registration.The method can be divided into two steps: The first step is to locate the layer of the anatomical points from the three-dimensional medical image;The second step is to locate the anatomical position of the anatomical point on the two-dimensional image corresponding to the layer of the anatomical point.The research contents of this paper are as follows:To locate the layer of the anatomical points,a method based on 3D convolutional neural network was proposed.The input of this method is a three-dimensional medical image(the data size is 16×515×512).3D convolutional neural network is used to extract the image features and continuously subsample the data of each layer through the pooling layer until the data size is reduced to 16×1×1 to generate a 16-dimensional vector and output it.Each value in this vector is the probability that the dissection point is at each level.In the network training process,the network output vector and the standard vector are made cross entropy loss and the loss is propagated back to complete the network training.The experimental results showed that the layer location accuracy of the four anatomical points reached more than 75% and the maximum location error was less than 1 layer.For the detection of anatomical points in two-dimensional images,an automatic detection method of dissection points in two-dimensional images based on point features and an automatic detection method of dissection points in two-dimensional images based on convolutional neural network are proposed.The point feature-based method uses the Haarlike feature of the image to generate the feature descriptor of a single pixel,and combines the multi-point joint description scheme to define the feature description vector of each pixel in the retrieval area of the test image.Then,these description vectors are matched with the standard vector one by one to detect the target point.The experimental results showed that the average error between the detection results of the four anatomical points in the twodimensional image based on Haar-like features and the results marked by doctors was between 2-4mm,which was smaller than the size of the anatomical points themselves.The automatic detection method of 2D image anatomical points based on convolutional neural network uses the hourglass network to extract image features and generate a predicted heat map of the anatomical points.The predicted heat map is converted into numerical coordinates of the anatomical points and output by DNST(Differentiable Spatial to Numerical Transform)layer.Because the computation in DSNT layer is differentiable,the network model realizes end-to-end training from input image to anatomical point coordinates.In order to enhance the training effect of the network,two methods of data enhancement and transfer learning are proposed.The experimental results show that the average detection error of the convolutional neural network-based method for detecting anatomical points in two-dimensional images is about 1mm,which is far better than the Haar-like feature based method.However,the convolutional neural network-based method requires a certain number of training sets to ensure the training effect.Finally,in order to realize the combination of the layer location method and the point location method,a software system of the algorithm is constructed.The system cascades and encapsulates the two methods to realize the automatic detection of anatomical points in 3D images.