A super-resolution algorithm to fuse orthogonal CT volumes using OrthoFusion

Abstract

OrthoFusion, an intuitive super-resolution algorithm, is presented in this study to enhance the spatial resolution of clinical CT volumes. The efficacy of OrthoFusion is evaluated, relative to high-resolution CT volumes (ground truth), by assessing image volume and derived bone morphological similarity, as well as its performance in specific applications in 2D-3D registration tasks. Results demonstrate that OrthoFusion significantly reduced segmentation time, while improving structural similarity of bone images and relative accuracy of derived bone model geometries. Moreover, it proved beneficial in the context of biplane videoradiography, enhancing the similarity of digitally reconstructed radiographs to radiographic images and improving the accuracy of relative bony kinematics. OrthoFusion's simplicity, ease of implementation, and generalizability make it a valuable tool for researchers and clinicians seeking high spatial resolution from existing clinical CT data. This study opens new avenues for retrospectively utilizing clinical images for research and advanced clinical purposes, while reducing the need for additional scans, mitigating associated costs and radiation exposure.

Keywords: Bone models; Computed tomography; Image Fusion; Spatial resolution enhancement; Super Resolution.

Fig. 1

OrthoFusion Super Resolution Algorithm. Orthogonal clinical volumes with low through-plane resolution (top row) are linearly interpolated onto an isotropic high-resolution 3-D grid (bottom row). A voxel-by-voxel average is computed to produce the super resolution volume (right).

Fig. 2

Study design. Four cases were extracted to evaluate the OrthoFusion algorithm – High Resolution (HR), Clinical (Clin), Resliced (RS), and Super Resolution (SR). The axial HR volume was considered the ‘ground truth’. First, HR CT volumes were acquired and reconstructed into 3 orthogonal volumes (axial, coronal, sagittal) with in-plane resolutions of ~ 0.2 × 0.2 mm and a through-plane resolutions of 0.6 mm (top row). Next, Clin volumes were simulated by down-sampling the HR volumes using Gaussian interpolation to a 3 mm through-plane resolution (middle row). The OrthoFusion algorithm (box) was applied, as depicted in Fig. 1. The RS case volume was derived from the axial Clin volume by interpolating back to a 0.6 mm through-plane resolution using Lanczos interpolation.

Fig. 3

The biplane videoradiography procedure for measuring intersegmental kinematics and performance measures across the pipeline. To capture the dynamic motion of individual vertebrae, biplane videoradiography is used in conjunction with corresponding bone model(s). This technique uses two synchronized radiographic units to simultaneously capture planar x-ray movies from two perspectives. In parallel, we create subject-specific bone models segmented from a CT scan. The process of shape-matching involves software that simulates a radiographic projection of the bone model – called a digitally reconstructed radiograph or DRR – onto the dynamic radiographic images. An optimization in each frame finds the closest match for the two projections to place and orient the vertebra in 3D space. The OrthoFusion Super Resolution approach was evaluated at each step along this process by quantifying key performance measures.

Fig. 4

Comparisons between the simulated Clinical (Clin), Resliced (RS), and Super Resolution (SR) volumes referenced to the “ground truth” High Resolution (HR) volume in (a) image similarity measures, (b) bone model morphological similarity, (c) shape-matching similarity, and (d) relative kinematics accuracy. Error bars represent standard deviation. Asterisk indicates a significant difference of p < 0.05 between groups. SSIM = Structural Similarity Index Measure, NCC = normalized cross-correlation coefficient.

Fig. 5

Summary Figure. Representative data illustrating the performance measures across the 4 case volumes, with the High Resolution case as the ground truth. Multiplanar Reconstructions emphasize the separation of the facet joints. Segmentation time was substantially less for the High Resolution and Super Resolution cases. Image similarity was improved with the Super Resolution approach, depicted from local SSIM maps where white indicates greater similarity. These improved performances over the Clinical and Resliced volumes yielded a greater morphological similarity and ultimately, for our applications purpose, the Super Resolution Digitally Reconstructed Radiograph (DRR) exhibited better performance for shape-matching similarity and relative kinematics accuracy.

Fig. 6

Qualitative evaluation of OrthoFusion applications. Top Row: MicroCT images of a mouse tibia. The 3 orthogonal “Clinical” CT volumes had a through-plane resolution of 0.075 mm and in-plane resolution of 0.005 × 0.005 mm. The Clinical images demonstrate representative slices in the plane of lowest spatial resolution, highlighting the loss of architectural detail. The Super Resolution reconstruction recaptures finer structural details, enhancing visualization of the trabecular architecture. Bottom Row: Segmented 3D models of a total knee arthroplasty (TKA) generated using different methods. The Clinical model exhibits significant blocky artifacts, hindering accurate segmentation and anatomical fidelity. The Super Resolution model mitigates these artifacts, producing a smoother and more accurate reconstruction.