Hippocampus segmentation in CT using deep learning: impact of MR versus CT-based training contours

Abstract

. Purpose: Hippocampus contouring for radiotherapy planning is performed on MR image data due to poor anatomical visibility on computed tomography (CT) data. Deep learning methods for direct CT hippocampus auto-segmentation exist, but use MR-based training contours. We investigate if these can be replaced by CT-based contours without loss in segmentation performance. This would remove the MR not only from inference but also from training. Approach: The hippocampus was contoured by medical experts on MR and CT data of 45 patients. Convolutional neural networks (CNNs) for hippocampus segmentation on CT were trained on CT-based or propagated MR-based contours. In both cases, their predictions were evaluated against the MR-based contours considered as the ground truth. Performance was measured using several metrics, including Dice score, surface distances, and contour Dice score. Bayesian dropout was used to estimate model uncertainty. Results: CNNs trained on propagated MR contours (median Dice 0.67) significantly outperform those trained on CT contours (0.59) and also experts contouring manually on CT (0.59). Differences between the latter two are not significant. Training on MR contours results in lower model uncertainty than training on CT contours. All contouring methods (manual or CNN) on CT perform significantly worse than a CNN segmenting the hippocampus directly on MR (median Dice 0.76). Additional data augmentation by rigid transformations improves the quantitative results but the difference remains significant. Conclusions: CT-based training contours for CT hippocampus segmentation cannot replace propagated MR-based contours without significant loss in performance. However, if MR-based contours are used, the resulting segmentations outperform experts in contouring the hippocampus on CT.

Deep learning in MRI‐guided radiation therapy: A systematic review

“…To account for MRI's high through‐plane resolution relative to its in‐plane resolution, 3D convolutional layers are often utilized to capture features not apparent in 2D convolution. However, 3D convolutions are computationally expensive, so numerous 2.5D architectures have been proposed 35–37 . In a 2.5D architecture, adjacent MRI slices are input as channels, and 2D convolutions are performed.…”

“…However, 3D convolutions are computationally expensive, so numerous 2.5D architectures have been proposed. 35 , 36 , 37 In a 2.5D architecture, adjacent MRI slices are input as channels, and 2D convolutions are performed. It is also common to see new papers forgo the 3D convolution to save resources for new computationally intense methods.…”