Implication of tumor morphology and MRI characteristics on the accuracy of automated versus human segmentation of GBM areas

An assessment scheme is proposed to evaluate GBM gross tumor core and T2-FLAIR hyper-intensity segmentations on preoperative multicentric MR images as a function of tumor morphology and MRI characteristics. 74 gross tumor core and T2-FLAIR hyper-intensity BraTS-Toolkit and DeepBraTumIA automatic segmentations, and 42 gross tumor core neurosurgeon manual segmentations were accordingly evaluated. Brats-Toolkit and DeepBraTumIA generally provide accurate segmentations, particularly for the most common round-shaped or well-demarked tumors, where: (1) gross tumor segmentation correctly includes necrosis and contrast enhanced tumor in 100% and 97.06% of cases (vs. 73.68% for manual segmentation) and wrongly includes healthy or non-tumor related tissues in 2.94% and 20.59% of cases (vs. 10.53% for manual segmentations); (2) T2-FLAIR hyper-intensity segmentations completely includes edema in 88.24% of cases for both software. MR image quality has little impact on the segmentation performance on these tumors. Conversely, on less common tumors with more complex tissue distribution and infiltrative behavior, manual segmentation works better than BraTS-Toolkit and DeepBraTumIA, and image quality has a larger impact on automatic segmentation performance. BraTS-Toolkit and DeepBraTumIA gross tumor segmentation properly includes necrosis and contrast enhanced areas in 50% and 37.50% of cases (vs. 66.67% for manual segmentation), all corresponding to higher image quality; T2-FLAIR hyper-intensity segmentation wrongly includes necrosis and contrast enhanced areas in 37.50% and 50% of cases.