OS9.2 Radiomics analysis of lower-grade gliomas, a POLA Network study. (6th September 2019)

• Record Type: Journal Article
• Authors: Younan, N; Douzane, H; Duran-Pena, A; Nichelli, L; Garcilazo, Y; Dehais, C; Ducray, F; Carpentier, C; Mokhtari, K; Figarella-Branger, D; Delattre, J; Idbaih, A; Alentorn, A
• Is Part Of: Neuro-oncology. Volume 21(2019)Supplement 3
• Page Start: iii18
• Page End: iii18
• Publication Date: 2019-09-06
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noz126.060
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: BACKGROUND: Lower-grade gliomas (LGG) are divided into three histo-molecular groups: i) IDH-wildtype, ii) IDH mutant and 1p19q intact and iii) IDH mutant and 1p19q co-deleted. The current classification has improved the clinical stratification and its reproducibility. However, LGGs are still associated with an important degree of clinical heterogeneity. We sought to analyze the cross-talk between the spatial distribution and the quantitative imaging features (radiomics) with the clinical evolution and their molecular background (radiogenomics). MATERIAL AND METHODS: We performed a retrospective multicentric study from 4 cohorts of high-grades gliomas (POLA Network, TCGA, REMBRANDT and LGG-1p19q), totaling 900 gliomas. We performed N4 and WhiteStripe imaging corrections to standardize MRI intensities. We used ITK-SNAP to obtain a mask of the different habitats of the tumor. Then we used PyRadiomics to obtain 2616 radiomic features per sample. We used plsRcox for fitting several Cox model in high-dimensional settings. We assessed the performance of the difference Cox model with the Harrel's concordance index. We used a Sparse Canonical Correlation analysis to analyze the spatial distribution of the tumors. RESULTS: Radiomics features allow identification in an unsupervised manner IDH-mutant gliomas with a median AUC of 0.96 [0.92–0.98]. Interestingly, in the analysis of survival, radiomics features provided additional information to clinical or genetics covariates and the model with only radiomics features obtained a C-Index of 0.78 [0.72–0.82]. In addition, survival model with the best performance in the prediction of overall survival was the one combining radiomics, clinics and genetics features with a C-Index 0.85 [0.82–0.92] and was validated in the other cohorts. The analysis of spatial distribution showed a very strong distribution of 1p19q co-deleted oligodendrogliomas in the frontal lobes. CONCLUSION: Radiomics features may provide additional relevant clinical information by improving the prognosis of LGG. Radiomics allow non-invasive prediction of the most relevant molecular alterations of LGG.