P15.03.A The value of 7 Tesla MR spectroscopic imaging for improved preoperative determination of the tumor grade and IDH status in gliomas: preliminary data. (5th September 2022)

• Record Type: Journal Article
• Authors: Hangel, G; Sharma, S; Cadrien, C; Furtner, J; Furtner, J; Roetzer-Pejrimovsky, T; Preusser, M; Bogner, W; Rössler, K; Trattnig, S; Widhalm, G
• Is Part Of: Neuro-oncology. Volume 24(2022)Supplement 2
• Page Start: ii84
• Page End: ii84
• Publication Date: 2022-09-05
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noac174.293
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Introduction: A new generation of MR spectroscopic imaging (MRSI) methods using 7T scanners have demonstrated the capability to resolve more neuro- and oncometabolites at higher resolutions than clinical routine MRSI. In a cohort of glioma patients, we explored the automated preoperative and noninvasive classification of IDH-mutation status and tumor grade based on 7T MRSI. Methods: This retrospective study included 36 patients (15 female) with histologically confirmed diffusely infiltrating glioma WHO grade 2-4 (9 grade 2, 9 grade 3 and 18 grade 4) and known IDH status (21 IDH1-mut, 15 IDH-wt) with an available 7T MRSI scan of sufficient data quality. The 3D MRSI scan had a 3.4 mm isotropic resolution and 15 minutes acquisition time. 12 spectral components were classified voxel-wise, including choline, glutamine and glycine. Within a tumor segmentation based on routine 3T imaging, we used a random forest algorithm for the voxel-wise classification of IDH mutation and grade (into low or high grade). Training used the leave-one-out cross validation method (i.e., for every patient data set, the other 35 datasets were used as training set) and feature selection out of the available combinations for metabolite ratios (e.g., glutamine to choline). The resulting voxel classifications were aggregated into a mean probability per patient that was the base for receiver-operator characteristic (ROC) curves both for grade and IDH status. Results: The classification algorithm obtained an area under the curve (AUC) for IDH determination of 0.85 (e.g., 75% sensitivity and 95% specificity). For grade determination, the AUC was 0.88 (e.g., 87% sensitivity and 89% specificity). In comparison, the AUC per voxel would have resulted in an AUC of 0.66 for both. Further, classification by individual metabolite ratios resulted in lower AUCs in all cases. Conclusions: According to our preliminary data, preoperative 7T MRSI is capable to determine the correct glioma grade and IDH status with high sensitivity and specificity by leveraging the extended metabolic panel width and voxel amount. By increasing this cohort in future, we intend to confirm our initial results and we also plan to extend classification to more molecular-pathological features (e.g., TERT). Thus, even a voxel-wise classification of tumor microenvironments could be attempted. Further improvements in 7T MRSI methodology such as absolute instead of relative quantification would also aid these attempts. In summary, 7T MRSI has shown its potential for improved preoperative characterization of gliomas.