Lung-CRNet: A convolutional recurrent neural network for lung 4DCT image registration.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence

Background:

• Deformable image registration (DIR) of lung 4DCT is crucial for clinical applications.
• Existing deep learning methods often overlook temporal continuity in deformation fields.
• There is a need for fast and accurate DIR methods that leverage the temporal nature of 4DCT.

Purpose of the Study:

• To propose a novel deep learning-based approach for lung 4DCT DIR that incorporates temporal information.
• To develop a fast and accurate method for registering lung 4DCT images, addressing limitations of pairwise registration.
• To improve the clinical utility of 4DCT by enhancing DIR accuracy and efficiency.

Main Methods:

• Lung-CRNet, a convolutional recurrent registration network, is presented for end-to-end lung 4DCT DIR.
• The method treats 4DCT DIR as a spatiotemporal sequence prediction problem, using ConvGRUs to capture temporal dynamics.
• The network is trained unsupervisedly with a spatial transformer layer and predicts displacement fields for each image pair.

Main Results:

• The Lung-CRNet achieved a mean target registration error of 1.56 ± 1.05 mm on the DIR-Lab dataset.
• Forward prediction computation time averaged less than 1 second per image pair.
• Performance was evaluated on a public 4DCT dataset and the DIR-Lab dataset.

Conclusions:

• Lung-CRNet demonstrates comparable accuracy and speed to current state-of-the-art deep learning DIR methods for lung 4DCT.
• The network's architecture is adaptable for other groupwise registration tasks involving multiple image alignments.
• The findings suggest Lung-CRNet is a promising tool for advancing lung 4DCT image analysis.