DART: DEFORMABLE ANATOMY-AWARE REGISTRATION TOOLKIT FOR LUNG CT REGISTRATION WITH KEYPOINTS SUPERVISION.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Lung CT registration is challenging due to lung deformability.
• Current deep learning models lack explicit anatomical knowledge.

Purpose of the Study:

• To develop an anatomy-aware deep learning model for lung CT registration.
• To improve keypoint-supervised registration accuracy using anatomical features.

Main Methods:

• Deformable Anatomy-aware Registration Toolkit (DART), a masked autoencoder (MAE)-based approach.
• Incorporates features from anatomical segmentation networks (lung, ribs, vertebrae, lobes, vessels, airways).
• Utilizes pretrained transformer encoder and patch embedding weights for initialization.

Main Results:

• DART outperforms baseline models (Voxelmorph, ViT-V-Net, MAE-TransRNet).
• Achieved relative improvement in target registration error for keypoints (17%, 13%, 9%).
• Achieved relative improvement in nodule center registration (27%, 10%, 4%).

Conclusions:

• DART enhances lung CT registration accuracy by integrating anatomical information.
• The MAE-based approach with anatomical priors offers a promising direction for medical image registration.