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

Area of Science:

• Medical Imaging
• Radiotherapy Planning
• Artificial Intelligence in Medicine

Background:

• Hippocampus segmentation for radiotherapy planning typically relies on MRI due to poor CT visibility.
• Deep learning models for CT segmentation often use MR-based training contours.
• Investigating CT-based contours aims to eliminate the need for MR data in both training and inference.

Purpose of the Study:

• To determine if CT-based contours can replace MR-based contours for training deep learning models for hippocampus auto-segmentation on CT.
• To assess the impact of using CT-based versus MR-based training contours on segmentation performance.
• To evaluate if deep learning models trained on CT contours can match or exceed manual expert contouring on CT.

Main Methods:

• Expert manual contouring of the hippocampus on MR and CT data from 45 patients.
• Training convolutional neural networks (CNNs) for CT hippocampus segmentation using either CT-based or propagated MR-based contours.
• Evaluating segmentation performance using Dice scores, surface distances, and contour Dice scores, with MR-based contours as ground truth.

Main Results:

• CNNs trained on propagated MR contours achieved a median Dice score of 0.67, significantly outperforming those trained on CT contours (0.59).
• CNNs trained on CT contours performed similarly to manual expert contouring on CT (0.59).
• CNNs trained on MR contours exhibited lower model uncertainty compared to those trained on CT contours.

Conclusions:

• CT-based training contours are insufficient for CT hippocampus segmentation without a significant performance decrease.
• Utilizing MR-based contours for training yields segmentations superior to manual expert contouring on CT.
• Eliminating MR data from training is not feasible without compromising segmentation accuracy for radiotherapy planning.