Research On Automatic Segmentation Method For Glioblastoma Multiforme In Multi-modal Resonance Imaging Using Machine Learning

Objective Glioblastoma multiforme(GBM)is an intracranial primary malignant tumour formed by astrocyte differentiation.It has the characteristics of low survival rate,high disability rate and high mortality.It is very harmful to patients.Whether it is the diagnosis of the disease,the presentation of the treatment plan or the prognosis,the GMB needs to be located and diagnosed.Multi-modal MR images contain a wealth of tissue structure information and are widely used in the diagnosis and treatment of GBM.Currently,the clinic relies mainly on radiologists to manually segment GBM using MR images.However,there are problems in the MR image that the gray level difference is not apparent,the surrounding is often edematous,the boundary is unclear,and the manual segmentation is complicated and cumbersome,and the repeatability is poor.Therefore,accurate manual segmentation of GBM presents significant challenges.The automatic method can avoid the false segmentation caused by human factors,and the result is relatively objective,which can significantly reduce the doctor's work intensity.Therefore,automatic segmentation of GBM multi-modal MR images has crucial clinical significance.Methods Aiming at the problem that most existing GBM multi-modal MR image segmentation algorithms do not realise fine segmentation of tumour regions,this paper proposes a random forest-based GBM multi-modal MR image segmentation method.First,the GBM three modal MR images are registered,and then the bias field correction is performed using the N4 ITK method;Secondly,after extracting the location feature,intensity feature,texture feature,context feature and symmetry feature of the MR image,the random forest classifier is applied to obtain the preliminary segmentation result;Finally,after removing the area smaller than 200 pixels,the 5×5 median filter is used to smooth the boundary of each area to obtain the random forest-based GBM segmentation result.On this basis,to further improve the accuracy of automatic segmentation for GBM,a GBM multi-modal MR image segmentation method based on three-dimensional region growing is proposed.First,the score map and the tumour area mask are calculated,and the tumour area information is obtained by using the tumour area mask,thereby analyzing the confidence and accuracy of the tumour area;Secondly,the seed point is selected for the three-dimensional region growing,and the high-scoring region is replaced with the corresponding region of the preliminary segmentation of the random forest model to generate new training data;Thirdly,The random forest model is trained with new training data,and the trained random forest model is used to classify each pixel based on the initial extracted 219 underlying features;Finally,after removing the area smaller than 200 pixels,the 5×5 median filter is used to smooth the boundary of each region to obtain the GBM segmentation result based on the three-dimensional region growing.Results In this paper,two kinds of segmentation methods are used to segment MR images into normal brain tissue regions,necrotic tumour regions,active tumour regions,T1 abnormal regions that remove necrotic tumour regions and active tumour regions,and FLAIR abnormal regions that do not contain T1 abnormal regions.The segmentation performance of the whole tumour region and the above four tumour sub-regions are evaluated by using the similarity coefficient,sensitivity and specificity.The experimental results show that the similarity coefficients of GBM multi-modal MR image segmentation methods based on the random forest are 0.869,0.748,0.857,0.768,0.681,respectively.The sensitivity is 0.851,0.739,0.872,0.745,0.667,respectively;The specificity is 0.9942,0.9949,0.9958,0.9937,and 0.9933.The GBM multi-modal MR image segmentation method based on three-dimensional region growth has similarity coefficients of 0.877,0.763,0.865,0.771,and 0.702,respectively.The sensitivity is 0.853,0.742,0.874,0.748,and 0.673,and the specificity is 0.9946,0.9965,0.9963,0.9941,and 0.9937,respectively.Conclusion It can be seen from the experimental results that the two GBM multi-modal MR image segmentation methods proposed in this paper can obtain better segmentation results.Through comparative analysis,the GBM multi-modal MR image segmentation method based on three-dimensional region growing shows superior performance in segmentation similarity,sensitivity and specificity of different tumour sub-regions.This paper provides an idea for automatically segmenting GBM multi-modal MR images,and has specific clinical reference value in the early diagnosis and surgical positioning of GBM.