Resolution enhancement of low-resolution skeletal images through reconstructing bone microstructure using topology optimization

Abstract

Although high-resolution skeletal images are essential for accurate bone strength assessment, the current HR imaging modalities have critical problems that remain to be solved such as high radiation doses, low signal-to-noise ratios, and long scan times. Resolution enhancement techniques, which have recently received much attention, have also been difficult to obtain acceptable image resolutions. Inspired by the self-optimizing capabilities of bone (i.e. reorienting the trabecula for maximum mechanical efficiency with minimum bone mass), this dissertation proposes the novel resolution enhancement of low-resolution skeletal images through reconstructing bone microstructure using topology optimization.

For the purpose, a fully automated segmentation using the patient-specific optimal thresholding and watershed algorithm firstly is proposed for extracting a target bone from medical scan data. Using the golden section method and load path algorithm, the proposed method first determines the patient-specific optimal threshold value that enables reliably separating two bones in a joint while removing cortical and trabecular bone in the femur at the minimum. This provides regional information on the femur. The watershed algorithm is then used to obtain boundary information on a target bone. The target bone can be extracted by merging the complementary information on a target image. For eight proximal femur CT images, compared with the manual segmentation and other segmentation methods, the proposed method offers a high accuracy in terms of the dice overlap coefficient and average surface distance within a fast timeframe in terms of processing time per slice. The proposed method also delivers structural behavior which is close to that of the manual segmentation with a small mean of average relative errors of the risk factor.

Secondly, estimation of local bone loads for the volume of interest is proposed in order to reduce excessive computational cost in the finite element analysis. The proposed method obtains physiological local load through partitioning and static condensation of a localized finite element model. The method is verified for the three VOI in a proximal femur in terms of force equilibrium, displacement field, and strain energy density (SED) distribution. The effect of the global load deviation on the local load estimation is also investigated by perturbing a hip joint contact force (HCF) in the femoral head. Deviation in force magnitude exhibits the greatest absolute changes in a SED distribution due to its own greatest deviation, whereas angular deviation perpendicular to a HCF provides the greatest relative change.

Thirdly, topology optimization-based bone microstructure reconstruction is proposed to enhance resolution of LR images. The proposed method conducts mesh refinement for resolution upscaling and then performs topology optimization with a constraint for the density deviation at multiresolution in order to preserve the subject-specific bone distribution data. The reconstructed trabecular bone includes the characteristic trabecular patterns and has morphometric indices that are in good agreement with the anatomical data in the literature. As for computational efficiency, the localization for the VOI reduces the number of FEs by 99%, compared with that of the full FE model. Compared with the previous single resolution density deviation constraint, the proposed multiresolution density deviation constraints enable at least 65% and 47% reductions in the number of iterations and computing time, respectively.

Finally, topology optimization-based bone microstructure reconstruction is validated. For the purpose, the proposed method used clinical QCT images and a micro-CT image. First, this study investigates whether the method can represent general bone characteristic pattern and age-related morphology difference. The bone strength difference between the LR input and the HR reconstructed images is investigated. Then, the method is thoroughly validated by comparing the reconstructed image and its reference image. The results demonstrate that the proposed method can reconstruct bone microstructure that has almost the same bone strength (maximum error is 1%) as well as bone morphology (maximum error 4% except SMI)