Evaluation of the HD-GLIO deep-learning algorithm for auto-2 matic segmentation of brain tumours during routine MRI treat-3 ment monitoring

Abstract

In the context of brain tumour response assessment, deep learning-based three-dimen-19 sional (3D) tumour segmentation has shown potential to enter the routine radiological workflow. 20 The purpose of the present study was to perform an external evaluation of a state-of-the-art deep-21 learning 3D brain tumour segmentation algorithm (HD-GLIO) on an independent cohort of consec-22 utive, post-operative patients. For sixty-six consecutive magnetic resonance imaging examinations, 23 we compared delineations of contrast-enhancing (CE) tumour lesions and non-enhancing T2/FLAIR 24 hyperintense abnormality (NE) lesions by the HD-GLIO algorithm and radiologists using Dice sim-25 ilarity coefficients (Dice). Volume agreement was assessed using concordance correlation coeffi-26 cients (CCCs) and Bland-Altman plots. The algorithm performed very well regarding segmentation 27 of NE volumes (median Dice = 0.79) and CE tumour volumes larger than 1.0 cm 3 (median Dice = 28 0.86). If considering all cases with CE tumour lesions, the performance dropped significantly (me-29 dian Dice = 0.40). Volume agreement was excellent with CCCs of 0.997 (CE tumour volumes) and 30 0.922 (NE volumes). The findings have implications for the application of the HD-GLIO algorithm 31 in the routine radiological workflow where small contrast-enhancing tumours will constitute a con-32 siderable share of the follow-up cases. Our study underlines that independent validations on clini-33 cal datasets is key to assert robustness of deep learning algorithms. 34