Session Item

Worse outcome of patients with glioblastoma (GBM) can be attributed to the heterogeneity of the tumor. Recent research has revealed that all GBM subtypes share the hallmark feature of aggressive invasion into the surrounding tissue that are not considered in the target volume. Then, identification of the different components of the tumor is of great importance to ensure an effective treatment. 1H MRI spectroscopic imaging (MRSI) is a non-invasive technique to obtain metabolic information. It is used for diagnosis, monitoring of response to treatment, dose escalation and for detection of mutated IDH gene status in gliomas. Every tissue can be classified based on its metabolic state. MRSI is able to identify the pathologic tissue with high accuracy where Choline (Cho), Creatine (Cr), N-acetylaspartate (NAA) and lactates (Lac) are the main altered metabolites.

The aim of this work is to 1) identify clusters of metabolic heterogeneity in GBM using large MRSI data and 2) investigate which clusters can predict the progression free survival (PFS).

MRSI data from 173 patients included in the prospective SPECTRO-GLIO Trial (Laprie et al, 2019 and 2021) were analysed. A total of 42 699 good quality spectra were acquired in pre-radiotherapy examination. Automatic post-processing of data was carried out with syngo.MR Spectro (VB40A; Siemens). Eight features were extracted: Cho/NAA, NAA/Cr, Cho/Cr, Lac/NAA and (%Cho, %NAA, %Cr, %Lac (ratio of each metabolite to the sum of all)).

The clustering was obtained via a mini-batch k-means algorithm, subsets of the input data are randomly sampled in each training iteration. Each mini-batch contains 1000 random spectra. The silhouette analysis was used to define the optimal number of clusters. The package used for these analyses were SuperML, factoextra and ClusterR. Cox model and Logrank test were used for PFS analysis. All analyses were conducted with R (4.0.2) and Stata software (16).

Five of clusters were identified as sharing similarities of metabolic information and are predictive of PFS. Clusters 2 and 4 revealed metabolic abnormalities (increase in Cho, stability or decrease in Cr, decrease in NAA and increase in Lac). The PFS was lower when cluster 4 or 2 are the dominant clusters in patient’s MRSI data. Table 1 shows that PFS were significantly different between clusters. Patients having foremost clusters 2 or 4 in their MRSI data have lower PFS.

The preliminary results showed that MRSI can be used to reveal heterogeneity of the GBM in pre-radiotherapy examination. Groups of spectra sharing similarities in metabolite information would reflect the different components of tissue in the MRSI acquired volume. Clusters with metabolic abnormalities representative of tumor proliferation and hypoxia, have been identified as predictor of progression free survival. Comparison of our clustering results with neuroradiologist segmentation of different tumor compartments on multimodal MRI data is ongoing.