Uncertainty-driven refinement of tumor-core segmentation using 3D-to-2D networks with label uncertainty

The BraTS dataset contains a mixture of high-grade and low-grade gliomas,\nwhich have a rather different appearance: previous studies have shown that\nperformance can be improved by separated training on low-grade gliomas (LGGs)\nand high-grade gliomas (HGGs), but in practice this information is not\navailable at test time to decide which model to use. By contrast with HGGs,\nLGGs often present no sharp boundary between the tumor core and the surrounding\nedema, but rather a gradual reduction of tumor-cell density.\n Utilizing our 3D-to-2D fully convolutional architecture, DeepSCAN, which\nranked highly in the 2019 BraTS challenge and was trained using an\nuncertainty-aware loss, we separate cases into those with a confidently\nsegmented core, and those with a vaguely segmented or missing core. Since by\nassumption every tumor has a core, we reduce the threshold for classification\nof core tissue in those cases where the core, as segmented by the classifier,\nis vaguely defined or missing.\n We then predict survival of high-grade glioma patients using a fusion of\nlinear regression and random forest classification, based on age, number of\ndistinct tumor components, and number of distinct tumor cores.\n We present results on the validation dataset of the Multimodal Brain Tumor\nSegmentation Challenge 2020 (segmentation and uncertainty challenge), and on\nthe testing set, where the method achieved 4th place in Segmentation, 1st place\nin uncertainty estimation, and 1st place in Survival prediction.\n