Automated Determination of the H3 K27-Altered Status in Spinal Cord Diffuse Midline Glioma by Radiomics Based on T2-Weighted MR Images

Study Design and Patients

This study was approved by the institutional review board of Beijing Tiantan Hospital, Capital Medical University, and written informed consent was obtained from all patients or their legal guardians.

A total of 128 patients with histopathologic diagnosis of diffuse glioma in the spinal cord from December 2013 to October 2019 were retrospectively enrolled. A total of 97 patients was finally included on the basis of the following inclusion criteria: 1) preoperative sagittal T2WI scans with adequate image quality; 2) available H3 K27-altered status. Exclusion criteria were the following: 1) a history of biopsy or an operation before MR imaging; 2) poor image quality (eg, obvious motion artifacts, distortion, and signal loss); 3) unknown H3 K27-altered status (Fig 1). From November 2019 to June 2022, thirty-nine patients with suspected SC-DMG were prospectively recruited for independent validation. Inclusion criteria were the following: 1) SC-DMG pathologically confirmed by biopsy or surgical resection; 2) no treatment history before MR imaging; 3) available H3 K27-altered status. Exclusion criteria were the following: 1) pathologic confirmation of other diseases; 2) poor image quality; 3) unknown H3 K27-altered status (Fig 1). MR images, including sagittal T2WI, T1WI, and contrast-enhanced (CE) CE-T1WI, as well as clinical variables, including patient demographics, treatment information, survival status, and detailed pathologic data, were obtained.

• Download figure
• Open in new tab
• Download powerpoint

FIG 1.

Flowchart of the included patients in this study.

After applying the inclusion and exclusion criteria, the final study cohort consisted of 97 patients in the retrospective cohort who were divided into a training set (n = 67, including 31 H3 K27-altered and 36 non-H3 K27-altered cases) and a test set (n = 30, including 14 H3 K27-altered and 16 non-H3 K27-altered cases) to build a H3 K27-altered predictive model. Then, an independent prospective cohort of 28 patients, including 16 H3 K27-altered and 12 non-H3 K27-altered cases, was enrolled to validate the predictive model. Further survival analysis was performed to confirm the clinical prognostic prediction value of the H3 K27-altered status of SC-DMG.

Neuropathology Analysis

All tumor histopathologic grading was evaluated following the 2016 WHO Classification using formalin-fixed/paraffin-embedded samples with H&E staining. The H3 K27-altered (H3 K27M mutation) status was assessed through immunohistochemistry using the corresponding specific antibody (ABE419; MilliporeSigma;1:800).

MR Imaging Acquisition

Spinal cord MR images were acquired using 6 MR imaging scanners: SIGNA HDe and Discovery 750 (GE Healthcare), Ingenia CX (Philips Healthcare), and Verio, Magnetom Trio, and Magnetom Prisma (Siemens). MR imaging protocols (T2WI, T1WI, and CE-T1WI) for intramedullary spinal cord lesions and parameters are found in the Online Supplemental Table 1.

MR Image Preprocessing

For all patients whose images were used with tumor segmentation, T2WI was first corrected for N4 bias using the SimpleITK python package (https://pypi.org/project/SimpleITK/). The image signal intensities were normalized by subtracting the mean value and dividing by the SD of the signal intensity.

Assessment of MR Imaging Characteristics

Conventional MR imaging features, including tumor location, length, edema, cystic/necrosis, contrast-enhancement status, and syringohydromyelia, were assessed by a senior neuroradiologist (Y.D., with >13 years of experience in neuroradiology) and a junior neuroradiologist (M.G., with >4 years of experience in neuroradiology) who were blinded to clinicopathologic data. Because there were high interrater agreements (K-values ranged from 0.71 to 1) on the MR imaging assessments, we presented the MR imaging assessments by the senior neuroradiologist in the main text and provided those by the junior neuroradiologist in the Online Supplemental Table 2. Tumor location was classified as cervical, cervicothoracic, thoracic, thoracolumbar, or lumbar vertebral, on the basis of the area involving the main component of the tumor. Tumor length was defined as the number of tumor-involved vertebrae. Edema was reflected by greater signal intensity than in the tumor noncontrast portion, lower signal intensity than in the CSF, and unclear boundaries on the T2WI.²⁰ The presentation of cystic or necrosis was well-defined as high-T2WI and low-T1WI signals essentially matching the CSF signals and nonenhanced wall. Compared with T1WI, the tumor CE pattern was classified as no enhancement (including mild enhancement) and obvious enhancement on CE-T1WI. Syringohydromyelia was defined as cystic dilation of the central portion of the spinal cord on T2WI without wall enhancement.^20,21

Radiomics Feature Extraction

Tumor segmentation on T2WI was automatically performed by a deep learning model developed by our research team (the mean Dice score of tumor segmentation was 0.852) and visually checked by a senior neuroradiologist (Y.D.) (8 of 97 patients in the retrospective cohort required additional manual review and correction).^22,23 Radiomics features of tumors were then extracted using Feature Analysis Explorer (FAE, Version 0.5.2; https://github.com/salan668/FAE) software. One hundred seven original radiomics features, including 14 shape-based features, 18 first-order features, 24 gray-level co-occurrence matrix features, 14 gray-level dependence matrix features, 16 gray-level run length matrix features, 16 gray-level size zone matrix features, and 5 neighboring gray tone difference matrix features, from T2WI were extracted.

Feature Processing and Selection

Using FAE, we adopted an up-sampling strategy to balance the H3 K27-altered and non-H3 K27-altered cases in the training set. Z score transformation was used for feature normalization. The Pearson correlation coefficient assessment was performed for feature-dimension reduction by randomly removing one of the paired features with high similarity (Pearson correlation coefficient > 0.99), followed by the Kruskal-Wallis method for selection of the optimal features.

Model Development

Logistic regression (LR) was used to determine the H3 K27-altered status by using the selected radiomics features. Five-fold cross-validation was performed for model parameter optimization in the training set. Then, we evaluated the performance of the model in the test set, by determining the classification accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC).

Sensitivity Analysis

Sensitivity analysis was performed to support the main findings using the LR radiomics model in the test set: 1) We compared the other 2 conventional MR imaging radiomic features, T1WI and CE-T1WI, to assess whether this additional information could improve the performance of the model. 2) To exclude potential bias in separating the training and test sets, we repeated the pipeline with 4 additional random splits of the training and test sets. 3) We included 4 classifiers, including support vector machine, least absolute shrinkage, and selection operator regression; linear discriminant analysis; and the Gaussian process for comparison to support the performance of the LR model. 4) We evaluated the predictive performance of combined original radiomics features and their variant features in comparison with the original radiomic features.

Survival Analysis

The overall survival (OS) of a subset of 79 patients was finally available, including 33 H3 K27-altered and 46 non-H3 K27-altered cases. The therapeutic regimen (eg, extent of tumor resection and postoperative treatment) was obtained from the medical records and follow-up visits. The extent of tumor resection was classified as total, subtotal, partial resection, or biopsy on the basis of medical records. Postoperative treatment included chemotherapy and radiation therapy.

Statistical Analysis

Statistical analysis was performed with SPSS (SPSS for Windows, Version 26.0; IBM) and the R statistical and computing software (Version 4.2.1; http://www.r-project.org/). Categoric variables were expressed as frequency and compared using the χ² test or Fisher exact test. Continuous variables were expressed as mean (SD) and compared using the Student t test or Mann-Whitney U test. P < .05 was considered statistically significant. Classification accuracy, sensitivity, specificity, PPV, NPV, and AUC were calculated to evaluate the performance of the model. OS was estimated by Kaplan-Meier survival curve analysis and the logrank test using the survminer package in R (https://cran.r-project.org/web/packages/survminer/index.html). Using the coxph package (https://www.rdocumentation.org/packages/survival/versions/3.5-7/topics/coxph), the univariate and multivariate Cox proportional hazards models were used to evaluate the model's performance with the WHO grade that integrating the H3 K27-altered status and histologic grade, and other clinical variables in predicting OS. P < .05 was considered statistically significant.