DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel at semantic image segmentation but require substantial labeled training data.
• Acquiring 3D medical image segmentations for supervised training is challenging and time-consuming.
• Classical methods explored joint segmentation and registration, inspiring a novel deep learning approach.

Purpose of the Study:

• To develop a deep learning framework that jointly learns image registration and segmentation networks.
• To leverage existing segmentations for registration supervision and generate segmentations when unavailable.
• To enhance segmentation network training through registration-based data augmentation.

Main Methods:

• A novel deep learning framework was proposed, integrating image registration and segmentation networks.
• The framework utilizes segmentations for weak supervision of registration and generates segmentations for augmentation.
• Experiments were conducted on 3D knee and brain MR images.

Main Results:

• The joint framework achieved simultaneous improvements in both segmentation and registration accuracy compared to independently trained networks.
• High-quality models were trained with significantly limited data, including a one-shot scenario.
• In the one-shot scenario, Dice scores increased by 2.7% for knee and 1.8% for brain images over unsupervised registration.

Conclusions:

• Joint learning of deep networks for image registration and segmentation is highly effective.
• This approach substantially reduces the dependency on large labeled datasets for medical image analysis.
• The framework demonstrates potential for training accurate models with minimal manual annotation.