|
Purpose: Cone-beam CT is increasingly used for 3D guidance in interventional radiology (IR), but long image acquisition time results in degradation from complex deformable motion of soft-tissue structures. Deformable motion compensation with multi-region autofocus optimization was shown to improve image quality. However, the high dimensionality and non-convexity of the optimization problem challenge its convergence. This work presents preliminary development and early results obtained from an automatic learning-based decision framework to obtain local estimates of basic properties of the deformable motion field, coupled to a preconditioning strategy to simplify the optimization. Methods: Deformable motion properties are estimated with a deep convolutional neural network (CNN) consisting of a concatenation of custom-designed residual blocks. The preliminary design provided an estimate of the local motion amplitude on an 8x8 grid covering an axial slice of a motion-contaminated CBCT volume. The decision framework is coupled to a preconditioning strategy that effectively favors more likely solutions through motion amplitude-driven spatially-varying regularization of the motion trajectory and spatially varying selection of the search range for the optimization problem. The network was trained on simulated data generated from publicly available CT datasets, including simple motion fields. Results: Predictions of local motion amplitude showed good agreement with the true values, with root mean squared error (RMSE) < 10 mm for the complete range of motion distributions explored (sufficient for the intended purpose of initialization). Combination of amplitude prediction with spatially varying regularization and search range setting resulted in improved motion compensation after 1000 iterations of the preconditioned multi-motion autofocus in an example case with complex deformable motion. Extensive validation in a large dataset of complex, multi-motion patterns is underway. Conclusion: The proposed approach shows promising initial results and the potential for automatic local motion estimation with learning-based methods. Pending ongoing development to extend this initial development, the method could simplify and accelerate complex deformable motion compensation with spatially varying preconditioning of the motion estimation.
|
