Click here to

Session: Machine Intelligence in Image Processing and Motion Correction II [Return to Session]

MUsculo-Skeleton-Aware Deep Learning-Based Deformable Registration for Head-And-Neck CT with a Relaxed Rigidity Constraint On Bony Structures

H Liu1,2*, E McKenzie3, Q Xu1,2, D Ruan1,2, K Sheng1,2, (1) UCLA, Los Angeles, CA, (2) UCLA School of Medicine, Los Angeles, CA, (3) Cedars-Sinai Medical Center, Los Angeles, CA


TH-F-BRC-5 (Thursday, 7/14/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Deep learning-based deformable image registration (DL-DIR) has demonstrated similar accuracy to time-consuming traditional methods in anatomical sites with relatively homogeneous tissue types. However, its performance is suboptimal in the presence of combined soft tissue deformation and musculoskeletal motion, e.g., the head-and-neck region. Here we proposed a MUsculo-Skeleton-Aware net (MUSA-net) to biomechanically guide DL-DIR using bony structures for intra-patient head-and-neck CT registration.

Methods: Bending energy is commonly used to regularize the smoothness of the deformation vector field. In this work, the loss was imposed with different weights for soft tissue and bony structures: a much higher weight for bones to encourage affine motion. This relaxed rigidity constraint is applicable for training on interpatient datasets as the bones are still allowed to scale and deform slightly to account for interpatient anatomical differences. We adopted Voxelmorph implementation with 3D-Unet architecture. An intra-patient dataset including 7 pairs of planning CT and PET attenuation correction CT was used for testing. Due to the scarcity of intra-patient CT pairs, our training used 392 interpatient head-and-neck CT from the TCIA database. Bony structures were segmented using existing deep learning methods. The proposed method was compared with vanilla Voxelmorph on target registration error (TRE) with manual landmarks.

Results: Visual inspection showed more realistic bone shape preservation using MUSA-net. The skeleton anchored the surrounding soft tissue with musculoskeletal motion to produce anatomically reasonable deformation. For the intra-patient test set, the proposed method achieved markedly reduced TRE (11.8 ± 6.2 mm), compared with Voxelmorph (20.4± 10.0 mm). B-spline registration using Elastix still achieved the lowest TRE (6.7 ± 1.8 mm) but took much longer time (70s vs 0.4s).

Conclusion: The proposed MUSA-net improved DL-DIR for head-and-neck CT. Future work will explore additional biomechanical properties, including explicit modeling of the joints, to provide better anatomical prior information for DL-DIR.


Registration, Deformation, Image Processing


IM/TH- Image Registration Techniques: Machine Learning

Contact Email