Deep learning-based automated detection and segmentation of bone and traumatic bone marrow lesions from MRI following an acute ACL tear.

INTRODUCTION: Traumatic bone marrow lesions (BML) are frequently identified on knee MRI scans in patients following an acute full-thickness, complete ACL tear. BMLs coincide with regions of elevated localized bone loss, and studies suggest these may act as a precursor to the development of post-traumatic osteoarthritis. This study addresses the labour-intensive manual assessment of BMLs by using a 3D U-Net for automated identification and segmentation from MRI scans. METHODS: A multi-task learning approach was used to segment both bone and BML from T2 fat-suppressed (FS) fast spin echo (FSE) MRI sequences for BML assessment. Training and testing utilized datasets from individuals with complete ACL tears, employing a five-fold cross-validation approach and pre-processing involved image intensity normalization and data augmentation. A post-processing algorithm was developed to improve segmentation and remove outliers. Training and testing datasets were acquired from different studies with similar imaging protocol to assess the model's performance robustness across different populations and acquisition conditions. RESULTS: The 3D U-Net model exhibited effectiveness in semantic segmentation, while post-processing enhanced segmentation accuracy and precision through morphological operations. The trained model with post-processing achieved a Dice similarity coefficient (DSC) of 0.75 ± 0.08 (mean ± std) and a precision of 0.87 ± 0.07 for BML segmentation on testing data. Additionally, the trained model with post-processing achieved a DSC of 0.93 ± 0.02 and a precision of 0.92 ± 0.02 for bone segmentation on testing data. This demonstrates the approach's high accuracy for capturing true positives and effectively minimizing false positives in the identification and segmentation of bone structures. CONCLUSION: Automated segmentation methods are a valuable tool for clinicians and researchers, streamlining the assessment of BMLs and allowing for longitudinal assessments. This study presents a model with promising clinical efficacy and provides a quantitative approach for bone-related pathology research and diagnostics.

Statistical investigation of interdependence of trabecular microarchitectural parameters from micro-computed tomography.

Standard microarchitectural analysis of bone using micro-computed tomography produces a large number of parameters that quantify the structure of the trabecular network. Analyses that perform statistical tests on many parameters are at elevated risk of making Type I errors. However, when multiple testing correction procedures are applied, the risk of Type II errors is elevated if the parameters being tested are strongly correlated. In this article, we argue that four commonly used trabecular microarchitectural parameters (thickness, separation, number, and bone volume fraction) are interdependent and describe only two independent properties of the trabecular network. We first derive theoretical relationships between the parameters based on their geometric definitions. Then, we analyze these relationships with an aggregated in vivo dataset with 2987 images from 1434 participants and a synthetically generated dataset with 144 images using principal component analysis (PCA) and linear regression analysis. With PCA, when trabecular thickness, separation, number, and bone volume fraction are combined, we find that 92 % to 97 % of the total variance in the data is explained by the first two principal components. With linear regressions, we find high coefficients of determination (0.827-0.994) and fitted coefficients within expected ranges. These findings suggest that to maximize statistical power in future studies, only two of trabecular thickness, separation, number and bone volume fraction should be used for statistical testing.

Establishing error bounds for internal calibration of quantitative computed tomography.

Opportunistic computed tomography (CT) scans, which can assess relevant osteoporotic bones of interest, offer a potential solution for identifying osteoporotic individuals. CT scans usually do not contain calibration phantoms, so internal calibration methods have been developed to create a voxel-specific density calibration that can be used in opportunistic CT. It remains a challenge, however, to account for potential sources of error in internal calibration, such as beam hardening or heterogeneous internal reference tissues. The purpose of this work was to introduce our internal calibration method that accounts for these variations and to estimate error bounds for the bone mineral density (BMD) measurements taken from internally calibrated scans. The error bounds are derived by incorporating a combination of a Monte Carlo simulation and standard error propagation into our previously established internal calibration method. A cohort of 138 clinical abdominal CT scans were calibrated for BMD assessment with a phantom placed in the field of view and used as the ground truth. Our modified internal calibration method provided error bounds on the same images and was tested to contain the ground truth phantom-calibrated BMD. This was repeated using 10 different internal reference tissue combinations to explore how error bounds are affected by the choice of internal tissue referents. We found that the tissue combination of air, skeletal muscle, and cortical bone yielded the most accurate BMD estimates while maintaining error bounds that were sufficiently conservative to account for sources of error such as beam hardening and heterogeneous tissue samples. The mean difference between the phantom BMD and the BMD resulting from the tissue combination of air, skeletal muscle and cortical bone was 2.12 mg/cc (0.06% BMD error) and 1.13 mg/cc (0.02 % BMD error) for the left and right femur, respectively. Providing error bounds for internal calibration provides a method to explore the influence of internal reference tissues and confidence for BMD estimates.

Automatic segmentation of trabecular and cortical compartments in HR-pQCT images using an embedding-predicting U-Net and morphological post-processing.

High-resolution peripheral quantitative computed tomography (HR-pQCT) is an emerging in vivo imaging modality for quantification of bone microarchitecture. However, extraction of quantitative microarchitectural parameters from HR-pQCT images requires an accurate segmentation of the image. The current standard protocol using semi-automated contouring for HR-pQCT image segmentation is laborious, introduces inter-operator biases into research data, and poses a barrier to streamlined clinical implementation. In this work, we propose and validate a fully automated algorithm for segmentation of HR-pQCT radius and tibia images. A multi-slice 2D U-Net produces initial segmentation predictions, which are post-processed via a sequence of traditional morphological image filters. The U-Net was trained on a large dataset containing 1822 images from 896 unique participants. Predicted segmentations were compared to reference segmentations on a disjoint dataset containing 386 images from 190 unique participants, and 156 pairs of repeated images were used to compare the precision of the novel and current protocols. The agreement of morphological parameters obtained using the predicted segmentation relative to the reference standard was excellent (R2 between 0.938 and > 0.999). Precision was significantly improved for several outputs, most notably cortical porosity. This novel and robust algorithm for automated segmentation will increase the feasibility of using HR-pQCT in research and clinical settings.

High performance multi-platform computing for large-scale image-based finite element modeling of bone.

BACKGROUND: Image-based finite element (FE) modeling of bone is a non-invasive method to estimate bone stiffness and strength. High-resolution imaging data as input allows for inclusion of bone microarchitecture but results in large amounts of data unsuitable for traditional FE solvers. Bone-specific mesh-free solvers have been developed over the past 20 years to improve on memory efficiency in simulated bone loading applications. The objective of this study was to provide linear performance benchmarking for a bone-specific, mesh-free solver (FAIM) using µCT and HR-pQCT image data on Mac, Linux, and Windows operating systems using both single- and multi-thread CPU and GPU processing. METHODS: The focus is on the linear gradient-descent solver using standardized uniaxial loading of bone models from µCT, and first- and second-generation HR-pQCT scans of the radius and tibia. Convergence, speedup, memory, and batch performance tests were completed using CPUs and GPUs on three laboratory-based systems with Windows, Linux, and Mac operating systems. RESULTS: Although varying by system and model size, time-per-iteration was as low as 0.03 s when an HR-pQCT-based radius model (6.45 million DOF) was solved with 3 GPUs. Strong scaling was achieved with GPU and CPU parallel processing, with strong parallel efficiencies when models were solved using 3 GPUs or ≤ 10 CPU threads. Errors in force, strain energy density, and Von Mises stress were as low as 0.1% when a convergence tolerance of 10-3 or smaller was used. CONCLUSION: The results of this study indicate that to maximize computational efficiency and minimize model solution times using FAIM software under the standardized tested conditions using µCT, XCT1 and XCT2 HR-pQCT image data, convergence tolerance set to 10-4, and 10 threads or 2 GPUs are sufficient for efficient solution times. Less strict convergence tolerances will improve solution times but will introduce more error in the outcome measures.

Diagnostic accuracy of a dual-energy computed tomography-based post-processing method for imaging bone marrow edema following an acute ligamentous knee injury.

OBJECTIVE: This study evaluated the ability of a custom dual-energy CT (DECT) post-processing material decomposition method to image bone marrow edema after acute knee injury. Using an independent validation cohort, the DECT method was compared to gold-standard, fluid-sensitive MRI. By including both quantitative voxel-by-voxel validation outcomes and semi-quantitative radiologist scoring-based assessment of diagnostic accuracy, we aimed to provide insight into the relationship between quantitative metrics and the clinical utility of imaging methods. MATERIALS AND METHODS: Images from 35 participants with acute anterior cruciate ligament injuries were analyzed. DECT material composition was applied to identify bone marrow edema, and the DECT result was quantitatively compared to gold-standard, registered fluid-sensitive MRI on a per-voxel basis. In addition, two blinded readers rated edema presence in both DECT and fluid-sensitive MR images for evaluation of diagnostic accuracy. RESULTS: Semi-quantitative assessment indicated sensitivity of 0.67 and 0.74 for the two readers, respectively, at the tibia and 0.55 and 0.57 at the femur, and specificity of 0.87 and 0.89 for the two readers at the tibia and 0.58 and 0.89 at the femur. Quantitative assessment of edema segmentation accuracy demonstrated mean dice coefficients of 0.40 and 0.16 at the tibia and femur, respectively. CONCLUSION: The custom post-processing-based DECT method showed similar diagnostic accuracy to a previous study that assessed edema associated with ligamentous knee injury using a CT manufacturer-provided, built-in edema imaging application. Quantitative outcome measures were more stringent than semi-quantitative scoring methods, accounting for the low mean dice coefficient scores.