A Research On AC/PC Localization Of MRI Brain Images Based On Deep Learning

The Magnetic Resonance Imaging(MRI)system is new type of medical device that uses magnetic fields to imaging human tissue.The use of magnetic resonance for brain science research is also a hot research topic.Anterior commissure and posterior commissure(AC/PC)are two important markers of the human brain,in magnetic resonance scan imaging,they are used to construct AC/PC reference coordinate systems for stereotactic neurosurgical planning,rapid pre-scanning of brain scans,pre-operative brain scans to detect conditions and develop targeted plans,and so on.Stereotactic radiosurgery is one of the best ways to treat neurological functional diseases such as Parkinson’s disease.Accurate identification of them is important for brain segmentation,registration,functional neurosurgery,pre-scan positioning,especially Talairach coordinate system conversion.In the past 20 years,the research on the anterior and posterior commissures of the brain requires manual intervention by professionals,but there are relatively few studies on the automatic localization of the anterior and posterior commissures of the human brain.This thesis aims to propose new method based on deep learning to locate the position of the AC/PC of the human brain MRI in a short period of time for non-radiologists.It is also to solve the existing problems of complex operation,weak robustness and high professional knowledge,this paper proposes two solutions.1、The AC/PC positioning based on 2D images.In most traditional methods,the positioning of the AC/PC is located under the sagittal plane.In the case of obtaining the sagittal plane,the computer vision YoLo framework is used to identify the partial region of the corpus callosum and the region of interest where the AC/PC is located.Finally,based on the region of interest,we design a dedicated positioning network by referring to VGG,Unet,FCN and other network models to locate the final position of the AC/PC.The test mean and standard deviation of the AC/PC are both within 1.600 mm and the test is performed at a speed of milliseconds.This method also verifies the feasibility of positioning the deep learning method.2、The AC/PC positioning based on 3D MRI images.This method ignores the intermediate process of sagittal plane positioning and directly performs end-to-end positioning.We use network model to extract spatial features,edge contours and implicit pixel information of 3D images.We design 3D models based on spatial and channel attention,GoogleNet’t inception module and long short term memory networks.The model of dimensional reduction of the mechanism reaches good result on high-resolution and low-resolution images,gets the fastest positioning speed of 0.004 s.