A generalized framework unifying image registration and respiratory motion models and incorporating image reconstruction, for partial image data or full images

Abstract

Surrogate-driven respiratory motion models relate the motion of the internal anatomy to easily acquired respiratory surrogate signals, such as the motion of the skin surface. They are usually built by first using image registration to determine the motion from a number of dynamic images, and then fitting a correspondence model relating the motion to the surrogate signals. In this paper we present a generalized framework that unifies the image registration and correspondence model fitting into a single optimization. This allows the use of 'partial' imaging data, such as individual slices, projections, or k-space data, where it would not be possible to determine the motion from an individual frame of data. Motion compensated image reconstruction can also be incorporated using an iterative approach, so that both the motion and a motion-free image can be estimated from the partial image data. The framework has been applied to real 4DCT, Cine CT, multi-slice CT, and multi-slice MR data, as well as simulated datasets from a computer phantom. This includes the use of a super-resolution reconstruction method for the multi-slice MR data. Good results were obtained for all datasets, including quantitative results for the 4DCT and phantom datasets where the ground truth motion was known or could be estimated.

Figure 1.

Typical approach for building a respiratory motion model. (1) Respiratory surrogate signals are acquired simultaneously with imaging data. (2) Image registration is used to determine the motion from the imaging data. (3) A correspondence model is fitted relating the motion to the surrogate signals.

Figure 2.

(a) ‘Lung-like’ 2D computer phantom. (b) The values of the correspondence model parameters used to animate the phantom: top-left  =  r_x,1, top-right  =  r_y,1, bottom-left  =  r_x,2, bottom-right  =  r_y,2.

Figure 3.

The MCIR produced from the real slab images dataset (left), a similar image formed without applying any motion correction (middle), and the end-exhalation 4DCT volume for the same subject (right). The dashed lines represent the height of the tumour and diaphragm in the end-exhalation 4DCT volume, and show that these are both lower in the MCIR. This is expected as the MCIR corresponds to the average position of the anatomy.

Figure 4.

Sagittal (top) and Coronal (bottom) views through the MCIR produced from the real thin slices dataset (left), a similar image formed without applying any motion correction (middle), and one of the fast helical scans from the same subject (right).

Figure 5.

One of the original dynamic images from the real thick slice dataset (left), and the corresponding simulated dynamic images from experiment 1 (middle) and experiment 2 (right).

Figure 6.

The MCIRs produced from the real thick slice dataset in experiment 1 using the averaging method (far-left) and experiment 2 using the super-resolution method (middle-left), and similar images formed using the averaging method (middle-right) and the super-resolution method (far-right) without applying motion correction.

Figure 7.

MCIRs generated from the phantom datasets assuming no motion (a)–(d) and using the fitted motion models to estimate the motion (e)–(h). The MCIRs were constructed by averaging the deformed dynamic images for the slab datasets (a) and (e), the thin slice dataset (b) and (f), and the thick slice dataset (c) and (g), and using the super-resolution method for the thick slice dataset (d) and (h).