Dlung: a new method for lung image registration

A research published in Journal of Shanghai Jiao Tong University (Science) has proposed a new method for lung image registration, named as Dlung. Dlung is an unsupervised few-shot learning-based diffeomorphic lung image registration, which can help construct respiratory motion models based on limited data with both high speed and high accuracy, offering a more efficient method for respiratory motion modeling.

Respiratory motion modeling is an essential technique in imaging technology for the analysis of thoracic organs such as lungs with respiratory motion. It offers important references for targeting tumors by radiotherapy while avoiding damage to normal tissues during treatment of lung cancer.

Lung image registration, the process of constructing a dense correspondence between lung image pairs, is critical for respiratory motion modeling. Among all the current methods for lung image registration, unsupervised learning-based methods have gained huge interest as they can compute the deformation without the requirement of supervision. However, there exist two drawbacks in the current unsupervised learning-based methods: one is that they are not able to handle problems with limited data; the other is that they lack diffeomorphic (topology-preserving) properties especially when large deformation exists in lung scans.

Aiming at these two problems, the researchers proposed the method Dlung which solves the problem of limited data via fine-tuning techniques and realizes diffeomorphic registration by the scaling and squaring method. Compared with baseline methods elastix, SyN and VoxelMorph, Dlung achieves the highest accuracy with diffeomorphic properties when applied in the registration of 4D images.

“Dlung constructs accurate and fast respiratory motion models with limited data,” explained Peizhi Chen, the first author of this research, “we believe that it has a wide application prospect in image-guided radiotherapy when treating lung cancer in the future.”