Automatic Acquisition Of Brainstem Region In MR Images

Brainstem is the basic physiological tissue for maintaining human heart rate,respiration,digestion,body temperature,sleep and other functions,and plays an extremely important role in maintaining human life.At the same time,the brain stem is also particularly vulnerable to direct or indirect external or internal stimulation,resulting in lesions or damage.It is one of the key steps in the diagnosis of brain stem lesions to obtain brain medical images,identify lesions or damaged areas and outline target areas.Therefore,the accurate and rapid automatic identification and segmentation of brainstem region in serial MR images is of great value for the diagnosis and treatment of diseases of head soft tissue and its attached physiological tissues.In order to solve the problem of low accuracy of brain stem segmentation in current MR images,which often requires manual intervention,this paper improves the classic Mask RCNN model on the basis of constructing the brainstem region data set in brain MR images,and compares other models,so as to obtain higher segmentation accuracy.This article was supported by the National Natural Science Foundation of China(No.61975248),the Natural Science Foundation of Guangdong Province(No.2018A0303130137),and the Science and Technology Project of Guangzhou(No.202007040004).The main work of this paper is as follows:1.A dataset of brain stem regions in MR brain images was constructed.The collected MR images of the brain were labeled,and whether the brain stem region was contained in the images and the boundaries of the brain stem region were delineated.Then professional physicians were invited to check the images,and finally a complete data set was created and divided.2.Design the research.On the established brainstem MR image data set,the clustering segmentation method based on pattern distance measurement and the deep learning method were used to perform brainstem region segmentation experiments respectively.By counting the segmentation results of different types of methods,we found a more suitable method for brainstem region segmentation task in brain MR sequence images.3.Improved Mask RCNN model.In view of the deficiency of the original Mask RCNN model in the recognition of brain stem region in brain MR sequence images,a pre-recognition network was added into the original model,that means the complete image information is used to prerecognize the brain stem region.By introducing more environmental information,the detection rate of brainstem region was improved.As the original Mask RCNN model will output segmentation results of multiple brainstem regions,the prior knowledge that a MR image has only one brainstem region at most is introduced to improve the output mode of the model and force the model to input only one brainstem region at most,making the model more suitable for brainstem segmentation.Designing a more appropriate loss function and the derivation and convergence of the function are proved in detail by mathematical process demonstration.The training method is optimized according to the improved model.By adopting the method of pre-training by stages and dynamically adjusting the weight of subobjective function blocks,the model can converge quickly,save training time and reduce training cost.The experimental results show that:(1)in the target region and background interference region has a very high similarity of the MR images,especially in the target area and there is no clear boundary between interference region,the segmentation result based on the fuzzy clustering method for DICE coefficient 76.89%,occurring simultaneously more than 62.45%,the original Mask RCNN model segmentation result is: the DICE coefficient of 93.68%,88.11%occurring simultaneously.(2)The improved mask-RCNN model proposed in this paper has a detection rate of 92.46%,a DICE of 94.33% and an intersection ratio of 89.27% for images containing brainstem region in the head sequence MR images.Compared with the original Mask RCNN model,the detection rate of brain stem region was increased by 37.19%.To sum up,the improved mask-RCNN model proposed in this paper can maintain the same segmentation accuracy and efficiency under the same training conditions,and can also effectively improve the detection rate of the brainstem region,realize automatic elimination of the images that do not contain the brainstem region in the serial MR images,and significantly reduce the workload of doctors’ manual image screening.