NI-13 The effectiveness and limitation of survival prediction in primary glioblastoma using machine learning-based texture analysis. (28th November 2020)

• Record Type: Journal Article
• Authors: Umehara, Toru; Kinoshita, Manabu; Sasaki, Takahiro; Arita, Hideyuki; Yoshioka, Ema; Shofuda, Tomoko; Kodama, Yoshinori; Hirayama, Ryuichi; Kijima, Noriyuki; Kagawa, Naoki; Okita, Yoshiko; Takano, Koji; Uda, Takehiro; Fukai, Junya; Sakamoto, Daisuke; Mori, Kanji; Kanemura, Yonehiro
• Is Part Of: Neuro-oncology advances. Volume 2(2020)Supplement 3
• Page Start: ii14
• Page End: ii14
• Publication Date: 2020-11-28
• Subjects: 616.99481
• DOI: 10.1093/noajnl/vdaa143.060
• Languages: English
• ISSNs: 2632-2498

Abstract

Abstract: Introduction: Clinical application of survival prediction of primary glioblastoma (pGBM) using preoperative images remains challenging due to a lack of robustness and standardization of the method. This research focused on validating a machine learning-based texture analysis model for this purpose using internal and external cohorts. Method: We included all cases of IDH wild-type pGBM available of preoperative MRI (T1WI, T2WI, and Gd-T1WI) from the databases of Kansai Molecular Diagnosis Network for CNS tumors (KN) and The Cancer Genome Atlas (TCGA). Of 242 cases from KN, we assigned 137 cases as a training dataset (D1), and the remaining 105 cases as an internal validation dataset (D2). Furthermore, we extracted 96 cases from TCGA as an external validation dataset (D3). Preoperative MRI scans were semi-quantitatively analyzed, leading to the acquisition of 489 texture features as explanatory variables. Dichotomous overall survival (OS) with a 16.6 months cutoff was regarded as the response variable (short/long OS). We employed Lasso regression for feature selection, and a survival prediction model constructed for D1 via cross-validation (M1) was applied to D2 and D3 to ensure the model robustness. Results: The population of predicted short OS by M1 significantly showed poorer prognosis in D2 (median OS 11.1 vs. 19.4 months; log-rank test, p=0.03), while there was no significant difference in D3 (median OS 14.2 vs. 11.9 months; p=0.61). In the comparative analysis using t-SNE, there was little variation in the feature distribution among three datasets. Conclusion: We were able to validate the prediction model in the internal but not in the external cohort. The presented result supports the use of machine learning-based texture analysis for survival prediction of pGBM in a localized population or country. However, further consideration is required to achieve a universal prediction model for pGBM, irrespective of regional difference.