Hi all! I’m not entirely sure how active this forum still being read/managed, but I wanted to try here first before moving to the fora of more recent challenges. In our current work, we are evaluating several segmentation models on in-house annotated data. For low-grade gliomas (LGG), we observed that models trained on BraTS data (on either the 2021 or 2023 pre-treatment glioma datasets) tend to perform relatively poorly. One of the main reasons appears to be differences in how necrotic regions are labeled. At our institute, we define necrosis quite strictly as regions that are hypointense on T1Gd, **surrounded by an enhancing rim**, and simultaneously hyperintense on T2. This definition is supported by literature from other groups. All remaining non-enhancing regions on T1Gd that are hyperintense on T2/FLAIR are labeled as non-enhancing tumor / infiltration and/or edema (we do not further distinguish between those entities on structural MRI alone). When reviewing the BraTS 2021 [paper](https://arxiv.org/pdf/2107.02314), we noted that the label previously referred to as necrosis _and_ non-enhancing tumor is now referred to simply as _just_ necrosis (NCR), defined as “the necrotic core of the tumor, the appearance of which is hypointense on T1Gd MRI.” While we fully appreciate the value of distinguishing necrosis from viable non-enhancing tumor, we do not really observe this distinguishment in the actual training data. In several LGG cases, relatively large regions are labeled as necrosis that, according to our criteria, would more likely be considered cystic or non-enhancing tumor rather than true necrosis. Some examples include BraTS2021_00012, BraTS2021_01533 & BraTS2021_01666. We would greatly appreciate some clarification on this point, mainy: Is the NCR/necrosis label in BraTS 2021 primarily based on hypointensity on T1Gd alone, without requiring additional features such as an enhancing rim or T2 characteristics? And if so, could you share some background on the rationale behind this choice? Thank you very much in advance for any insights, and for all the work that has gone into creating and maintaining the BraTS datasets.