Click here to

Session: AI Applications in Image Guided Adaptive Radiation Therapy [Return to Session]

Scale-Adaptive Convolutional Neural Network for Deformable Image Registration of Lung 4DCT

Y Sang*, D Ruan, Department of Radiation Oncology and Department of Bioengineering, University of California, Los Angeles, CA


TH-E-TRACK 4-4 (Thursday, 7/29/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: Deformable image registration between phases in lung 4DCT is essential for understanding respiratory motion pattern and facilitates temporally conformal radiotherapy delivery. Multi-resolution hierarchical strategy is typically used in conventional optimization-based registration to capture varying magnitudes of motions while avoiding undesirable local minima. A rough concept of the scale is captured in deep networks by the reception field of kernels, and it has been realized to be both desirable and challenging to capture convolutions of different scales simultaneously in registration networks. In this study, we propose a registration network that is conscious of and self-adaptive to motions of various scales to improve registration performance.

Methods: Dilated inception modules (DIMs) are proposed to incorporate receptive fields of different sizes in a computationally efficient way. Scale adaptive modules (SAMs) are proposed to guide and adjust shallow features using convolutional kernels with spatially adaptive dilation rate learned from deep features. DIMs and SAMs are integrated into a registration network which takes a U-net structure. The network was trained in an unsupervised setting with 15 4DCT scans and then tested using 10 scans in the DIR-Lab dataset, each with 300 annotated landmark pairs.

Results: The method achieved a (2.52±2.96) mm target registration error on the anatomical landmarks, better than a conventional B-spline method in SimpleElastix and networks without DIM or SAM, with statistical significance. The adaptive dilation rate in SAM corresponded well to the spatially and temporally varying motion scale. The average registration time was 0.42 s for images with size 256×256×96.

Conclusion: The introduction and integration of DIMs and SAMs address the spatio-temporal scale heterogeneity problem in an efficient and self-adaptive way and can be an efficacious alternative to the conventional multi-resolution strategy.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by a research grant from Varian Medical Systems, Inc.



    Registration, CT, Respiration


    IM/TH- Image Registration Techniques: Machine Learning

    Contact Email