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

Abstract

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.

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

“…Similar methods have been used elsewhere for downgrading CT images 35,36 . The pairs of synthetic ground‐truth and blurry clinical CT bone samples provide then an infinite synthetic training dataset with feasibly still enough realism to train data‐driven methods of clinical applications, such as a convolutional neural network for tissue‐specific noise removal 2 . Since the synthetic bone pairs are not expected to be perfectly realistic, the method requires still some additional pairs of real HRpQCT and clinical CT scans to be used for fine‐tuning or validation.…”

“…Such a method may gain insight in the microstructure of osteopenic or osteoporotic bone, or help to analyze the performance of new microstructural parameters. It can also be used as a basis for an in silico simulation of clinical CT scans 1 to provide pairs of synthetic ground‐truth and clinical CT that can be used then to train data‐driven methods for CT noise suppression 2 . Additionally, when the latent space of the generative network is smooth and adequately well formed, one might be able to develop a simulation that transforms a bone sample into another just as it would degenerate under osteoporosis.…”

Effective Spatial Resolution of Photon Counting CT for Imaging of Trabecular Structures is Superior to Conventional Clinical CT and Similar to High Resolution Peripheral CT

“…This allowed the repetition of microstructural measurements on different CT scanners. After scanning all 12 vertebrae with XCT, we chose a subset of 5 vertebra phantoms according to their bone mineral density, trabecular separation, and size as prototypes of the entire distribution for further tests; details of the selection process are published elsewhere 3 . XCT scans were obtained with 60 kVp, 170 mAs, and isotropic voxel size of 82 μm and reconstructed with the sharp kernel.…”

“…Radial and Axial Minimum Resolvable Object Sizes, Radial and Axial Effective Spatial Resolutions, and Effective Volumetric Spatial Resolutions for All Examined Scanners and Kernels at High (10% MTF) and Low (5% MTF, Values in Brackets) ContrastEffective volumetric resolution is a statistical measure of the number of points that can be resolved in a cube of 1 mm 3 ; lp/cm indicates line pairs per centimeter; resels/ mm 3 , resolution elements per mm3 . *Due to downsampling, the axial and radial effective resolution of XCT (10% MTF) is limited to 30.5 lp/cm (164 μm).…”

Whole-body magnetic resonance imaging (WBMRI) versus whole-body computed tomography (WBCT) for myeloma imaging and staging

“…A 3D convolutional neural network was recently introduced with specific loss functions for QCT noise reduction to compute microstructural parameters such as tissue mineral density and bone volume ratio. This method allows the assessment of the 3D bone microstructure and has the potential to detect rarefaction of the trabecular network due to osteoporosis or other bone diseases [ 43 ]. It may further discriminate patients with and without vertebral fractures or indicate early stages of osteolytic processes that are not yet visible [ 44 ].…”