Abstract
Gliomas represent a heterogenous group of primary brain tumors with overlapping imaging phenotypes. Treatment typically includes surgery and/or chemoradiation, however this varies based on the overall lesion and clinical presentation. This heterogeneity in both lesion characteristics and management strategies contributes to a lack of reliable findings when evaluating treatment outcomes with conventional MRI. The overlapping imaging features of radiation necrosis and tumor progression post-treatment can be particularly challenging for radiologists. We present a dataset of 203 glioma patients with 594 post-treatment timepoints of relevant clinical history and routine T1, T1 postcontrast, T2, and FLAIR weighted MR sequences. Preprocessing of the images follow a standardized pipeline with automatic deep-learning based segmentations for each tumor component i.e. enhancing tumor, non-enhancing necrotic core, surrounding non-enhancing FLAIR signal hyperintensity, and resection cavity. The automatic segmentations were manually validated and refined by neuroradiologists to get the ground truth labels. Our contribution of this robust dataset to an open-source repository aims to contribute to the development of AI models to improve evaluation of treatment outcomes.