Improving Early Post-Operative Glioblastoma Segmentation With Semi-Supervised Deep Learning

Improving automatic segmentation of glioblastoma on early post-operative MRI is key to study the effect of resection volumes on patient outcomes. We curate a dataset of over 700 MRI examinations, of which 87 include annotations, and train a supervised and a semi-supervised deep-learning model. Semi-supervision improves the segmentation of the high-intensity FLAIR signal with 3% to a Dice score of 0.83 (p=0.031), while the segmentation of the enhancing tumor increases with 9% to 0.55 (p=0.056). However, enhancing tumor segmentations show high variability, possibly due to imperfect annotations. Segmentation of enhancing tumor on early post-operative MRI remains a challenging task.