Research On Key Techniques Of Radiomics-Based Glioma Grading

Glioma is the most common malignant tumor in the brain.It is also the most difficult-to-cure tumor among intracranial tumors.In the Central Nervous System Tumor Classification Guidelines issued by the World Health Organization,gliomas are divided into grades Ⅱ-Ⅳ.Grades Ⅰ and Ⅱ are low-grade gliomas,and grades Ⅲ and Ⅳare high-grade gliomas.High-grade gliomas are more aggressive and have worse prognosis than low-grade gliomas.The accurate preoperative grading of glioma is important for patient treatment and prognostic evaluation.Magnetic resonance(MR)imaging is the preferred imaging method for patients with glioma.Qualitative evaluation is clinically provided through manual reading of radiographers.In recent years,with the development of artificial intelligence and the proposal of radiomics,the quantitative analysis of images by radiomics to predict glioma grading provides new ideas for clinical research.In recent years,although radiomics have achieved staged results in glioma grading,there are still the following problems to break through:(1)Labeled and multi-center glioma datasets are scarce.Most researches are still based on public datasets or smallscale datasets from a single source,lacking the necessary external verification and it is difficult to find more general rules.(2)The existing multi-center image datasets have not been standardized.The data usually comes from different equipment imaging models of multiple medical institutions,different scanning parameters and different imaging protocols are set,and the images are represented by different resolutions,different contrasts,and brightness.It is bound to have a potential impact on the accuracy of the radiomics model,and conventional image preprocessing methods are difficult to effectively alleviate the variability of multi-center data.(3)The manual segmentation of region of interest(ROI)by radiologists has a huge workload and low efficiency.At present,manual segmentation of ROI is still regarded as the"gold standard",but manual segmentation is highly repetitive and the segmentation task is arduous,which can easily cause the radiologist to have subjective deviations during segmentation.In view of the several key issues in the grading of glioma in the above radiomics,the main basic and innovative work of this article is as follows:1)The multicenter glioma database used in this study was constructed.The glioma public dataset BraTS2017 was integrated,retrospective cases of glioma in the First Affiliated Hospital of Zhengzhou University were collected and sorted out,the internal dataset GI2019 was constructed,and the public dataset and internal dataset were combined to enlarge the multicenter dataset.In accordance with the standards for public dataset,the internal dataset was subject to data type standardization and privacy elimination processing.In addition,in view of the differences in image contrast and brightness in all data,histogram specification method was proposed to alleviate the potential re-effects of image contrast and brightness differences on the glioma grading.Experiments showed that the standardized images processed by the histogram specification algorithm could improve the accuracy of glioma grading and provide new ideas for the study of image standardization.2)In order to solve the problem of subjective bias and extra interaction caused by manual selection of mapping baseline image by radiologists in histogram specification method,this paper proposes the Histogram Specification of Parameters ControllableGrid Search(HSPC-GS).HSPC-GS uses a Gaussian probability density function that can control the two parameters of contrast and brightness to construct the histogram of the baseline image,and then generates a standardized image through the histogram specification.This algorithm can overcome the subjective deviation of the radiologists.In order to solve the optimal standardized image,the glioma grading of radiomics is designated as the grid search task,and the AUC is defined as the evaluation index.The grid search adjusts the parameters according to the AUC level to realize the automatic solution of the optimal standardized image.Experiments show that the data set processed by the HSPC-GS algorithm has excellent stability.Compared with the histogram specification method,the classification accuracy is further improved.3)Aiming at the problem that the original image information extracted by the HSPC-GS algorithm is not sufficient,this paper proposes Pixel Correlation-based Histogram Specification of Parameters Controllable and-Grid Search(PCHSPC-GS).HSPC-GS algorithm only describes the distribution of gray level of image,and does not consider the correlation between pixels and neighboring pixels.In fact,the shape and nature of tumor are mainly described by the change of gray value of pixels.The feature of gray value correlation between pixels and neighboring pixels can be extracted to describe image information more specifically.Therefore,PCHSPC-GS uses two-dimensional histogram to model for the multi-center original images,which can extract the feature of pixel correlation.Secondly,the two-dimensional Gaussian function is constructed correspondingly,and the controllable parameters are preserved.PCHSPC-GS also uses grid search strategy,which defines the classification of gliomas and the prediction of IDH genotype as grid search task.The optimal combination of parameters is found through grid search,while the optimal standardized image set is obtained.The results show that the standardized data set processed by PCHSPC-GS algorithm also keeps excellent stability,and the accuracy rate is improved.4)Aiming at the problem of huge workload and low efficiency of manual ROI segmentation by radiologists.In this paper,we propose a pooling fusion algorithm for MR images of gliomas,called Slices Pooling(SP).SP algorithm generates fused images by fusing the gray values of the same position on different MR slices.The traditional labeling method is changed from slice by slice to single slice.In addition,we developed two radiomics models for glioma grading and IDH state prediction,and verified the effectiveness of the fusion image in clinical tasks through the models.Experimental results show that SP algorithm can significantly reduce the time of radiologists in ROI segmentation,and the fused image can still maintain excellent performance.The glioma classification model based on fusion image can maintain good diagnostic accuracy,while the accuracy of IDH state prediction model is slightly decreased.