Bone-GAN: Generation of virtual bone microstructure of high resolution peripheral quantitative computed tomography

BACKGROUND

Data-driven development of medical biomarkers of bone requires a large amount of image data but physical measurements are generally too restricted in size and quality to perform a robust training.

PURPOSE

This study aims to provide a reliable in silico method for the generation of realistic bone microstructure with defined microarchitectural properties. Synthetic bone samples may improve training of neural networks and serve for the development of new diagnostic parameters of bone architecture and mineralization.

METHODS

One hundred-fifty cadaveric lumbar vertebrae from 48 different male human spines were scanned with a high resolution peripheral quantitative CT. After prepocessing the scans, we extracted 10,795 purely spongeous bone patches, each with a side length of 32 voxels (5 mm) and isotropic voxel size of 164 μm. We trained a volumetric generative adversarial network (GAN) in a progressive manner to create synthetic microstructural bone samples. We then added a style transfer technique to allow the generation of synthetic samples with defined microstructure and gestalt by simultaneously optimizing two entangled loss functions. Reliability testing was performed by comparing real and synthetic bone samples on 10 well-understood microstructural parameters.

RESULTS

The method was able to create synthetic bone samples with visual and quantitative properties that effectively matched with the real samples. The GAN contained a well-formed latent space allowing to smoothly morph bone samples by their microstructural parameters, visual appearance or both. Optimum performance has been obtained for bone samples with voxel size 32 × 32 × 32, but also samples of size 64 × 64 × 64 could be synthesized.

CONCLUSIONS

Our two-step-approach combines a parameter-agnostic GAN with a parameter-specific style transfer technique. It allows to generate an unlimited anonymous database of microstructural bone samples with sufficient realism to be used for the development of new data-driven methods of bone-biomarkers. Particularly, the style transfer technique can generate datasets of bone samples with specific conditions to simulate certain bone pathologies.

Distribution of bone voids in the thoracolumbar spine in Chinese adults with and without osteoporosis: A cross-sectional multi-center study based on 464 vertebrae

Bone void is a novel intuitive morphological indicator to assess bone quality but its use in vertebrae has not been described. This cross-sectional and multi-center study aimed to investigate the distribution of bone voids in the thoracolumbar spine in Chinese adults based on quantitative computed tomography (QCT). A bone void was defined as a trabecular net region with extremely low bone mineral density (BMD) (<40 mg/cm³), detected by an algorithm based on phantom-less technology. A total of 464 vertebrae from 152 patients (51.8 ± 13.4 years old) were included. The vertebral trabecular bone was divided into eight sections based on the middle sagittal, coronal, and horizontal planes. Bone void of the whole vertebra and each section were compared between healthy, osteopenia, and osteoporosis groups and between spine levels. Receiver operator characteristic (ROC) curves were plotted and optimum cutoff points of void volume between the groups were obtained. The total void volumes of the whole vertebra were 124.3 ± 221.5 mm³, 1256.7 ± 928.7 mm^3, and 5624.6 ± 3217.7 mm³ in healthy, osteopenia, and osteoporosis groups, respectively. The detection rate of vertebrae with bone voids was higher and the normalized void volume was larger in the lumbar than in thoracic vertebrae. L3 presented the largest void (2165.0 ± 3396.0 mm³), while T12 had the smallest void (448.9 ± 699.4 mm³). The bone void was mainly located in the superior-posterior-right section (40.8 %). Additionally, bone void correlated positively with age and increased rapidly after 55 years. The most significant void volume increase was found in the inferior-anterior-right section whereas the least increase was found in the inferior-posterior-left section with aging. The cutoff points were 345.1 mm³ between healthy and osteopenia groups (sensitivity = 0.923, specificity = 0.932) and 1693.4 mm³ between osteopenia and osteoporosis groups (sensitivity = 1.000, specificity = 0.897). In conclusion, this study demonstrated the bone void distribution in vertebrae using clinical QCT data. The findings provide a new perspective for the description of bone quality and showed that bone void could guide clinical practice such as osteoporosis screening.

Dose-efficient assessment of trabecular microstructure using ultra-high-resolution photon-counting CT

Photon-counting (PC) detectors for clinical computed tomography (CT) may offer improved imaging capabilities compared to conventional energy-integrating (EI) detectors, e.g. superior spatial resolution and detective efficiency. We here investigate if PCCT can reduce the administered dose in examinations aimed at quantifying trabecular bone microstructure. Five human vertebral bodies were scanned three times in an abdomen phantom (QRM, Germany) using an experimental dual-source CT (Somatom CounT, Siemens Healthineers, Germany) housing an EI detector (0.60 mm pixel size at the iso-center) and a PC detector (0.25 mm pixel size). A tube voltage of 120 kV was used. Tube current-time product for EICT was 355 mAs (23.8 mGy CTDI_{32 cm}). Dose-matched UHR-PCCT (UHRdm, 23.8 mGy) and noise-matched acquisitions (UHRnm, 10.5 mGy) were performed and reconstructed to a voxel size of 0.156 mm using a sharp kernel. Measurements of bone mineral density (BMD) and trabecular separation (Tb.Sp) and Tb.Sp percentiles reflecting the different scales of the trabecular interspacing were performed and compared to a gold-standard measurement using a peripheral CT device (XtremeCT, SCANCO Medical, Switzerland) with an isotropic voxel size of 0.082 mm and 6.6 mGy CTDI_{10 cm}. The image noise was quantified and the relative error with respect to the gold-standard along with the agreement between CT protocols using Lin's concordance correlation coefficient (r_CCC) were calculated. The Mean ± StdDev of the measured image noise levels in EICT was 109.6 ± 3.9 HU. UHRdm acquisitions (same dose as EICT) showed a significantly lower noise level of 78.6 ± 4.6 HU (p = 0.0122). UHRnm (44% dose of EICT) showed a noise level of 115.8 ± 3.7 HU, very similar to EICT at the same spatial resolution. For BMD the overall Mean ± StdDev for EI, UHRdm and UHRnm were 114.8 ± 28.6 mgHA/cm³, 121.6 ± 28.8 mgHA/cm³ and 121.5 ± 28.6 mgHA/cm³, respectively, compared to 123.1 ± 25.5 mgHA/cm³ for XtremeCT. For Tb.Sp these values were 1.86 ± 0.54 mm, 1.80 ± 0.56 mm and 1.84 ± 0.52 mm, respectively, compared to 1.66 ± 0.48 mm for XtremeCT. The ranking of the vertebrae with regard to Tb.Sp data was maintained throughout all Tb.Sp percentiles and among the CT protocols and the gold-standard. The agreement between protocols was very good for all comparisons: UHRnm vs. EICT (BMD r_CCC = 0.97; Tb.Sp r_CCC = 0.998), UHRnm vs. UHRdm (BMD r_CCC = 0.998; Tb.Sp r_CCC = 0.993) and UHRdm vs. EICT (BMD r_CCC = 0.97; Tb.Sp r_CCC = 0.991). Consequently, the relative RMS-errors from linear regressions against the gold-standard for EICT, UHRdm and UHRnm were very similar for BMD (7.1%, 5.2% and 5.4%) and for Tb.Sp (3.3%, 3.3% and 2.9%), with a much lower radiation dose for UHRnm. Short-term reproducibility for BMD measurements was similar and below 0.2% for all protocols, but for Tb.Sp showed better results for UHR (about 1/3 of the level for EICT). In conclusion, CT with UHR-PC detectors demonstrated lower image noise and better reproducibility for assessments of bone microstructure at similar dose levels. For UHRnm, radiation exposure levels could be reduced by 56% without deterioration of performance levels in the assessment of bone mineral density and bone microstructure.

Noise reduction using novel loss functions to compute tissue mineral density and trabecular bone volume fraction on low resolution QCT

Micro-structural parameters of the thoracic or lumbar spine generally carry insufficient accuracy and precision for clinical in vivo studies when assessed on quantitative computed tomography (QCT). We propose a 3D convolutional neural network with specific loss functions for QCT noise reduction to compute micro-structural parameters such as tissue mineral density (TMD) and bone volume ratio (BV/TV) with significantly higher accuracy than using no or standard noise reduction filters. The vertebra-phantom study contained high resolution peripheral and clinical CT scans with simulated in vivo CT noise and nine repetitions of three different tube currents (100, 250 and 360 mAs). Five-fold cross validation was performed on 20466 purely spongy pairs of noisy and ground-truth patches. Comparison of training and test errors revealed high robustness against over-fitting. While not showing effects for the assessment of BMD and voxel-wise densities, the filter improved thoroughly the computation of TMD and BV/TV with respect to the unfiltered data. Root-mean-square and accuracy errors of low resolution TMD and BV/TV decreased to less than 17% of the initial values. Furthermore filtered low resolution scans revealed still more TMD- and BV/TV-relevant information than high resolution CT scans, either unfiltered or filtered with two state-of-the-art standard denoising methods. The proposed architecture is threshold and rotational invariant, applicable on a wide range of image resolutions at once, and likely serves for an accurate computation of further micro-structural parameters. Furthermore, it is less prone for over-fitting than neural networks that compute structural parameters directly. In conclusion, the method is potentially important for the diagnosis of osteoporosis and other bone diseases since it allows to assess relevant 3D micro-structural information from standard low exposure CT protocols such as 100 mAs and 120 kVp.

Three-dimensional multifractal analysis of trabecular bone under clinical computed tomography

PURPOSE

An adequate understanding of bone structural properties is critical for predicting fragility conditions caused by diseases such as osteoporosis, and in gauging the success of fracture prevention treatments. In this work we aim to develop multiresolution image analysis techniques to extrapolate high-resolution images predictive power to images taken in clinical conditions.

METHODS

We performed multifractal analysis (MFA) on a set of 17 ex vivo human vertebrae clinical CT scans. The vertebrae failure loads (FFailure) were experimentally measured. We combined bone mineral density (BMD) with different multifractal dimensions, and BMD with multiresolution statistics (e.g., skewness, kurtosis) of MFA curves, to obtain linear models to predict FFailure. Furthermore we obtained short- and long-term precisions from simulated in vivo scans, using a clinical CT scanner. Ground-truth data - high-resolution images - were obtained with a High-Resolution Peripheral Quantitative Computed Tomography (HRpQCT) scanner.

RESULTS

At the same level of detail, BMD combined with traditional multifractal descriptors (Lipschitz-Hölder exponents), and BMD with monofractal features showed similar prediction powers in predicting FFailure (87%, adj. R² ). However, at different levels of details, the prediction power of BMD with multifractal features raises to 92% (adj. R2) of FFailure. Our main finding is that a simpler but slightly less accurate model, combining BMD and the skewness of the resulting multifractal curves, predicts 90% (adj. R2) of FFailure.

CONCLUSIONS

Compared to monofractal and standard bone measures, multifractal analysis captured key insights in the conditions leading to FFailure. Instead of raw multifractal descriptors, the statistics of multifractal curves can be used in several other contexts, facilitating further research.