Correlation of T1 to T2 Weighted Signal Intensity Ratio with T1 and T2 Relaxation Time and IDH Mutation Status in Glioma



Patient cohort

The study was approved by the institutional review boards of Asahikawa Medical University Hospital (Approval No. 21041) and the International Cancer Institute, Osaka (No. 1612065191). The requirement for written informed consent has been waived for data collected retrospectively. Patients in the prospectively recruited cohorts received a detailed explanation and written informed consent was obtained from each patient or their family. The survey was conducted in accordance with all relevant local guidelines and regulations, and adhered to the principles of the Declaration of Helsinki.

We first reanalyzed previously published T1 and T2 relaxation data of LrGGs histologically16 and compared these T1 and T2 relaxometry data with rT1/T2 (the reanalyzed cohort, Fig. 1). T1 and T2 relaxometry was performed using MP2RAGE and multi-echo T2-weighted images of nine patients with glioma. Thus, we were able to access the T1 and T2 relaxometries as well as the rT1/T2 data of these nine patients, and included the data of eight of them in our analyses, after excluding a patient with a mutant tumor. K27M.16.

Figure 1

Overall study cohort. First, the correlation of rT1/T2 with T1 and T2 relaxation time was investigated by reanalyzing the raw data from ref.16. Then, the study was conducted in two stages, as an exploratory cohort study followed by a validation cohort study to investigate the correlation of rT1/T2 and HDI histologically confirmed LrGG mutation status. IDHmt, IDH mutant; IDHwt, wild-type IDH; AMUH, Asahikawa Medical University Hospital; OICI, Osaka International Cancer Institute; TCIA/TCGA, Cancer Imaging Archives/Cancer Genome Atlas.

We then prepared a new set of three cohorts and conducted a two-stage study (Fig. 1). The first cohort (exploratory cohort) included 25 histologically and molecularly confirmed LrGGs (HDIwt: 8, HDImt: 17) treated at Asahikawa Medical University Hospital (AMUH). The second and third cohorts were used as validation cohorts. Validation Cohort 1 included 29 patients (IDHwt: 13, HDImt: 16) from Osaka International Cancer Institute (OICI) and Validation Cohort 2 included 101 patients (IDHwt: 19, HDImt: 82) from Cancer Imaging Archive ( TCIA)/Cancer Genome Atlas (TCGA) Low-Grade Gliomas Collection Dataset, accessed 1 February 202020.21. Validation cohort 1 can be considered as a “domestic” validation cohort and validation cohort 2 as an “international” validation cohort.

Pathologic diagnosis was based on the 2016 WHO classification of tumors of the central nervous system22. IDHwt tumors in the current cohorts could not be fully characterized according to the 2021 WHO classification system because molecular analyzes such as TER promoter mutation, EGFR gene amplification and +7/−10 chromosome copy number alterations had not been achieved. The inclusion criterion was the availability of T1WI and T2WI. Excluded were patients with failed image co-registration or insufficient or atypical images (eg, with tumor hemorrhage). Table S1 provides detailed information on the three cohorts.

Genetic analysis

Two laboratories performed genetic analyzes of glioma tissues: the Department of Pathology, Asahikawa Medical University, Asahikawa, Japan, for the exploratory cohort; and the Department of Biomedical Research and Innovation, Institute for Clinical Research, Osaka National Hospital, Osaka, Japan, for validation cohort 1. Immunobiological detection of HDI1 mutation was performed for the exploratory cohort, and Sanger sequencing was performed to detect hotspot mutations of HDI1/2 (codon 132 of HDI1 and codon 172 of HDI2) for validation cohort 13. The HDI the mutation status of tumors in the TCIA/TCGA dataset was obtained from the report by Ceccarelli et al.23.

Reconstruction of T1w/T2w (rT1/T2) images

Table S1 lists the MRI acquisition parameters in detail. Most of the images of the exploratory cohort were acquired with General Electric 3 T scanners (Chicago, Illinois, USA), and those of the validation cohort 1 with Siemens 3 T scanners (Erlangen, Germany). Images from the TCIA cohort (the validation cohort 2) were acquired by 1.5 and 3 T scanners from various MRI vendors. The T1WI and T2WI in Digital Imaging and Communication in Medicine (DICOM) formats were converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format using Mango software (version 4.0.1; University of Texas Health Science Center,, accessed 6 March 2022). We used in-house imaging software integrating an algorithm for reconstructing rT1/T2 images from T1WI and T2WI24. The algorithm and MATLAB codes for calculating rT1/T2 can be obtained as an open source toolkit for SPM12 developed by Ganzetti et al. ( March 6, 2022)18. Details of the reconstruction analysis have already been reported by Ganzetti et al.18. Reconstruction was performed by first applying bias field correction to the original T1WI and T2WI using SPM12 (, accessed 6 March 2022). Intensity histograms were then adjusted based on intensities extracted from non-brain tissues such as cerebrospinal fluid, bone, and soft tissue. Finally, the processed T2WI was co-registered and split by the processed T1WI to produce an rT1/T2 image using the NIfTI “scanner-anatomical” coordinate system (Fig. 2).

Figure 2
Figure 2

Workflow on the whole study. T1WI to T2WI (rT1/T2) signal intensity ratio images were calculated from T1- and T2-weighted images, after image normalization via bias field correction and l ‘histogram. Voxels of interest (VOI) were defined manually based on the high intensity area of ​​pathological lesions on T2-weighted images, followed by the mean rT1/T2 measurement in the VOI.

Multi-echo MP2RAGE and T2WI T1 and T2 relaxometry

Imaging was performed on a 3 T MR scanner (Prisma; Siemens Healthcare, Erlangen, Germany). T1 relaxation was performed by converting MP2RAGE images into T1 relaxation time maps. T2 relaxation was performed by converting multi-echo T2-weighted images into T1 relaxation time maps. In both cases, relaxometry was performed via Bayesian inference modeling (Olea Nova+; Canon Medical Systems, Tochigi, Japan). Other technical details have been reported previously16.

Segmentation of voxels of interest (VOI) and calculation of the rT1/T2 average

Author 1, with 6 years of experience in neurosurgery, performed manual lesion segmentation by designing voxels of interest (VOI) in ITK-SNAP software (version 3.8.0,, accessed 6 March 2022). VOIs were designed on T2WIs with visual identification of pathologically high-intensity areas, avoiding ambiguous and vaguely abnormal lesions as much as possible (Fig. S1). The last author, with 22 years of experience in neurosurgery, then evaluated the VOIs and either confirmed their position or requested modification (which happened for five VOIs) (Table S1). The Dice similarity coefficient ranged from 0.52 to 0.81 for these VOIs (TCIA-00067, TCIA-00071, TCIA-00110, TCIA-00111, TCIA-00113). This procedure was performed on T2WI using the NIfTI “general affine transformation” coordinate system.

Each rT1/T2 image on the NIfTI “scanner-anatomical” coordinate system was then co-registered with T2WI on the NIfTI “general affine transformation” coordinate system using Volume Imaging in Neurological Research, Co-Registration and included ROI ( VINCI; Max Planck Institute for Neurological Research Cologne, Germany,, accessed 6 March 2022), to ensure that further analysis will be performed using the NIfTI “general affine transformation” coordinate system (Fig. S2). Three-dimensional VOIs were then applied to the rT1/T2 images for the calculation of the mean rT1/T2 (mrT1/T2) in the VOIs (Fig. 2). VOIs were also applied to T1 and T2 relaxation time maps when these data were available.

Image Feature Extraction

Image features were extracted from T1WI and T2WI using the method described previously25. T1WI and T2WI were converted to 256-level grayscale images after clipping the upper 0.1% of the signal. This procedure was not performed for rT1/T2 images, as these are quantitative in nature. First-order texture features were calculated based on the 256-level gray scale histograms in the VOIs of the T1WI, T2WI and rT1/T2 images. Second-order texture features were measured in gray-level co-occurrence matrix (GLGM) and gray-level run-length matrix (GLRLM) analyses. A total of 49 imaging features were extracted from each image (Table S2). Details of the extracted imaging features have been provided previously25. Texture features were used only to compare image features among cohorts and were not used to predict HDI state of mutation. Such multiparametric variables would require a large set of training data to establish a reliable prediction model.

Statistical analysis and t-Distributed Stochastic Neighbor Embedding (t-SNE)

Statistical analysis was performed using Prism 9 for macOS (GraphPad Software, San Diego, CA, USA). The relationship between mrT1/T2 and HDI mutation status was investigated by Mann-Whitney U test and receiver operating characteristic (ROC) curve analysis. A p-a value below 0.05 was considered significant. t-Distributed Stochastic Neighbor Embedding (t-SNE) analysis was used to investigate the difference in MRI qualities and characteristics between the three cohorts. Version 0.15 of the Rtsne package for R with default settings was used for this analysis (Tables S3, S4 and S5)26.

Source link


Comments are closed.