Patients selection
A retrospective study was performed at the XX Hospital of XX University. The patients from May 2015 to May 2018 were retrospectively analyzed with the following criteria: 1. Incipient brain glioma patients with WHOII-III grading (IDH1 immuno-histochemical results confirmed) 2. High-quality MRI imaging included T1WI, T2WI, DWI (ADC image), Flair and post-contrast T1WI series. 3.All the patients had not received the radiotherapy and chemotherapy treatment or given antitumor drug before surgery. Images could not be used for analysis were excluded. As a result, a total of 146 patients (77 males and 69 females) were enrolled in the study. Among them, 60 were IDH1 mutants and 86 were IDH1 wild type. The data were divided into training group and validation group according to the ratio of 7:3. The training group included 102 patients (42 IDH1 mutants and 60 wild type), and the validation group included 44 patients (18 IDH1 mutants and 26 wild type).
MRI imaging acquisition
All patients were examined following the same imaging acquisition protocol on a 1.5T MRI system (Philip Achieva Dutch) with a sixteen-channel head matrix coil. The conventional MR imaging protocols consisted of the following sequences: axial T1 weighted gradient-echo imaging (TR500ms;TE12ms), axial T2-weighted spin-echo imaging (TR5000ms;TE110ms), axial fluid-attenuated inversion recovery imaging (TR8400ms;TE120ms), and 3 orthogonal plane contrast-enhanced gradient-echo T1-weighted imaging scans (TR250ms;TE2.48ms) acquired with a dosage of 0.1mmol/kg. A high-pressure syringe was used for rapid injection from the elbow vein. The flow rate was 2ml/s.The section thickness (5mm), intersection gap (1 mm), and 240x240 mm FOV,256x256 matrix were uniform in all sequences. DWI was performed in the axial plane with a spin-echo echoplanar sequence before injection of contrast material. The imaging parameters used were as follows:TR4200ms;TE72.3ms, slice thickness:5mm, intersection gap: 1mm, FOV:220x220mm. The b-values were 0 and 1000s/mm2 with diffusion gradient sencoded in the three orthogonal directions to generate 3 sets of diffusion-weighted images. Processing of the ADC image was generated automatically by the MR imaging system.
Immunohistochemistry staining and Histopathological Dignosis
Immunohistochemistry was performed on 5-μm-thick sections from paraffin-embedded tumour specimens of all evaluated patients. IDH1R132H Mouse anti-Human Monoclonal Antibody (concentrated 1:50) was purchased from Fuzhou Maixin Biotechnology Co., Ltd. The samples for IDH1 mutation detection were repaired by EDTA antigen repair solution (pH=9.0) and citric acid buffer solution (pH=6.0). The samples were fixed by 4% neutral polyformaldehyde. EnVision two-step immunohistochemical staining was used for routine sampling, dehy-dration and paraffin-embedded tissue sections. The instrument used was LEICABOND-MAX automatic immunohistochemical staining instrument (LEICA, Germany). Normal autopsy brain tissue was used as negative control, and the positive section of known glioma IDH1 mutation was used as positive control. The staining results were determined by selecting 5 standard visual fields under 40-fold objective microscope. The expression of IDH1 mutant protein was localized in cytoplasm. Any number of brown cells appeared in the visual field of tumors, and the stained cells were regarded as positive mutations.
Statistics analysis :
Statistics analysis consists of three specific steps: 1. Screening statistical differences variables of general clinical data and conventional morphological features. 2. Texture lables(including peritumoral edema and tumor parenchyma)acquisition and extraction. 3. Establishing and evaluating the preoperative diagnosis model of IDH1 gene type expression. Specific steps are as follows:
1.Screening statistical differences variables of general clinical data and conventional morphological features:The general clinical data of the two groups were recorded, including gender and age. Morphological scores based on VASARI were performed in all patients. All the data were analyzed by the R language (v.3.3.2 http://www.R-project.org) software. Shapiro-Wilk test was used to test the normality of quantitative data. The data conforming to normal distribution were expressed by (`x±s). Those who did not conform to the normal distribution were expressed as median ± quartile intervals. The quantitative data between IDH1 mutant and wild type groups were compared by independent sample t-test. Two samples were extracted according to the characteristic set of VASARI (Visual Accessible Rembrandt Images) of glioma. The Rembrandt Visual Accessible Images (VASARI) feature set is organized and planned by the radiologists of the Cancer Genome Atlas (TCGA) team (Location) based on MRI sequences T1WI, T2WI, FLAIR, DWI and enhanced T1WI conventional quantitative imaging features of gliomas which quantified the 30 qualitative descriptive features (included gender, age, location of tumors, maximum diameter,peritumoral edema, intratumoral cystic degeneration, enhancement value of tumors,percentage of necrosis and extent of resection, etc.) and standardize the extraction process of quantitative features. In our experiment,we excluded the 26th, 27th, and 28th features which were not consistent with the purpose to predict IDH1 gene expression type before operation,but are more indicative of post-operative survival rate. Ultimately,27 morphological features of tumors before the operation were evaluated. All the quantitative imaging features (Feature 1 to Feature 30, excluded the 26th, 27th and 28th ) were carried out by using single-factor analysis. Pearson's Chi-Square test was used for ranking data, and t- test for classifying variables was used for two independent samples. Independent predictors of IDH1 mutation were screened by multivariate Logistic regression (P<0.05).
2.Texture lables(including peritumoral edema and tumor parenchyma)acquisition and extraction:(1)IBEX (v1.0 beta Dr. Laurence E, Court's Core Lab MD Anderson Cancer Center Houston TX77030) volume texture analysis software was performed based on MATLAB 2014 b platform. Volumetric ADC image texture parameters were drawn from the images of tumor parenchyma and peritumoral edema manually with ROI (Region of Interest) separately, and all volume texture parameters of both parenchyma and peritumoral edema were extracted, Figure1.The volume texture parameters of tumor parenchyma include six categories: Gradient Orient Histogram, Gray Level Cooccurrence Matrix, Intensity Direct, Intensity Histogram, Neighbor Intensity Difference and Shape with a total of 1733 texture features. The boundary of peritumoral edema refers to the FLAIR sequence. Because of subtracting the intermediate tumor entities, the second-order texture parameters cannot be evaluated. Therefore, we only choose histogram-based texture parameters to analyze. There were 50 texture parameters for edmea volumetric ADC image at last, concluding: Mean Absolute Deviation, Variance, Skewness, Kurtosis, percentiles (Percentile.01%, Perc.10%, Perc.50%, Perc.90%, Perc.99%). The definition of texture parameters is explained as follows: 1.Degree of skewness: A class of statistics describing the symmetry of variable distribution, i.e. the degree of asymmetry of distribution relative to mean. Both positive and negative skewness show that the distribution is asymmetric. The larger the absolute value of skewness, the greater the skewness and the deviation from the normal distribution.(2)Kurtosis: statistics describing the shape steepness of a variable distribution, which reflects the degree of flatness or steepness compared with a normal distribution. The positive value is the steep peak, indicating that it is steeper than a normal distribution, while the negative value is the flat peak, indicating that it is flatter than a normal distribution. 3.Percentile: Describe the percentage of subjects observed below that percentile, reflecting small changes in the region of interest, closely related to the heterogeneity of lesions, the heterogeneity of lesions is large, the difference between high percentiles is greater, on the contrary, the difference between low percentiles is the smallest(22).
Texture analysis generates a large amount of information, and multivariate image analysis (MIA) could be a valuable tool for understanding and interpreting the data. The most essential part of the MIA procedure is principal component analysis (PCA).
Principal component analysis (PCA) was used to screen out texture labels from thousands of texture parameters which can indicate invisible pretreatment prognostic information. PCA is frequently used for reducing multi-dimensional data into a few principal segments and screening out latent variables that contain the majority of the variation in the data(23,24) .
All the extracted image texture parameters have seemed as variables in PCA statistics analysis. Variables with correlation coefficients greater than 0.999 were screened out by correlation analyses. The main components a cumulative contribution rate of more than 85% were selected, and the most influential feature vectors were found as the most important factors.