Deep learning-based brain tumor segmentation on limited sequences of magnetic resonance imaging.

Abstract

2054 Background: Deep learning algorithms trained to segment brain tumors from magnetic resonance imaging (MRI) perform well in ideal conditions supplied by curated datasets. An example is the MICCAI BraTS 2018 dataset, which provides a set of four MRI sequences (T1, T1 post-gadolinium contrast-enhanced, T2, FLAIR) on 285 samples, along with ground truth segmentations annotated by experts. In real-world settings, however, it is not uncommon for patients to undergo MRI with a limited number of image sequence acquisitions. We examined the effect of restricting the imaging sequence set when training a deep learning model for brain tumor segmentation. Our goal was to identify a limited subset that still achieved acceptably high performance. Methods: In our experiments, we used the convolutional neural network-based U-Net architecture. Instead of the standard BraTS task, we focused on sub-segmenting specific areas of brain tumors: the active, enhancing tumor (AT) and the tumor core (TC; AT plus the cystic/necrotic core). We trained the architecture on 285 samples of the BraTS 2018 dataset using one or both of two sequences: T1 contrast-enhanced (T1CE) and FLAIR. Using each training model, we predicted segmentation masks given a volumetric MR image of a brain tumor on 66 samples in the held-out validation dataset. We submitted the predicted segmentations to the Center for Biomedical Image Computing and Analytics (CBICA) imaging portal for evaluation, which reports prediction accuracy by returning dice scores on each prediction. Results: As shown in the Table, the U-Net model using T1CE alone yielded high performances, with median dice scores of 0.84461 and 0.88267, when segmenting AT and TC, respectively. On the other hand, using FLAIR alone generated lower performances on the same tasks of segmenting the AT (0.25996) and the TC regions (0.51153). Combining T1CE and FLAIR provided no additional performance improvement compared with T1CE used as a standalone sequence. Conclusions: We achieved high performances in two brain tumor sub-segmentation tasks using a limited number of MRI sequences in training a deep-learning algorithm. T1CE alone produced high performances on both TC and AT segmentations. Given the ubiquity of the T1CE sequence, the ability to achieve high performance tumor segmentations in the face of limited image sequence availability is critical when applying our algorithm in real-world clinical or research settings. [Table: see text]