Can Brats Toolkit Brain Tumor Segmentation Algorithms Be Relevant When Dealing With Post-Operative Mri In Clinical Practice?

Abstract BACKGROUND Segmentation of brain tumors is crucial for radiotherapy plans and treatment outcome evaluation. As manual segmentation (MS) is a time-consuming task, many algorithms to automate the process were proposed over the past decades. BraTS toolkit (BTK) offers a solution for automatic brain tumor segmentation in 3 steps: a Preprocessor for image conversion and registration, a Segmentor generating segmentations from 4 deep-learning based algorithms and a Fusionator combining the results. As most algorithms published, BTK was trained on pre-operative data. Yet, as surgery is the first treatment of glioblastomas, most MRIs in clinical practice are post-operative images. This study aimed to assess whether segmentation of post-operative brain tumors could benefit from an initial automatic segmentation (AS) using BTK. MATERIAL AND METHODS MRI dataset from the multicenter, STERIMGLI clinical trial provided 92 series from 25 patients with a unifocal recurrence of glioblastoma. AS were generated using BTK Preprocessor and Segmentor. Out of the 4 algorithms output, the best AS was selected after visual appraisal. AS contained 3 labels: T1w contrast enhanced tumor (ET), flair edema (ED) and non-enhanced tumor (NET). AS were then reviewed by a radiation oncologist and a neuroradiologist to produce the MS. ET and ED were corrected; surgical cavity (SC) was also segmented, either from the NET label or de novo. Dice-score, Hausdorff Distance (HD) and Average Hausdorff Distance (AHD) were used to quantify the similarity between AS and MS for each label. RESULTS AS succeeded to label 89.3% (82/92), 100% (92/92) and 85.8% (79/92) of ET, ED and NET respectively. Among the 4 algorithms in BTK, Zyx_2019 produced 36% of AS, mic-dkfz 32%, xfeng 18%, lfb_rwth 14%. Mean Dice-scores of 75.8%, 94.8% were found for ET and ED respectively. Mean HD and AHD were 25.2mm (±39.7), 2.8mm (±10.9) for ET; 14.9mm (±25.8), 0.9mm (±5.8) for ED. Concerning SC, Dice-scores were <0.1 for 49% (39/79), >0.6 for 30% (42/79). CONCLUSION BraTS Toolkit was trained to segment necrosis as the label NET. Still, it detected the surgical cavity and saved time for the MS in 51% of post-operative cases. Even though BTK was designed to segment pre-operative brain tumors, similarity metrics show that minimal or no manual corrections are necessary most of the time when used to segment ET and ED on post-operative MRI acquired in clinical routine. The development of a unique AS tool for pre and post-operative images would be useful in clinical practice.