Simultaneous automated ascertainment of prevalent vertebral fracture and abdominal aortic calcification in clinical practice: role in fracture risk assessment.

Whether simultaneous automated ascertainments of prevalent vertebral fracture (auto-PVFx) and abdominal aortic calcification (auto-AAC) on vertebral fracture assessment (VFA) lateral spine bone density (BMD) images jointly predict incident fractures in routine clinical practice is unclear. We estimated the independent associations of auto-PVFx and auto-AAC primarily with incident major osteoporotic and secondarily with incident hip and any clinical fractures in 11 013 individuals (mean [SD] age 75.8 [6.8] years, 93.3% female) who had a BMD test combined with VFA between March 2010 and December 2017. Auto-PVFx and auto-AAC were ascertained using convolutional neural networks (CNNs). Proportional hazards models were used to estimate the associations of auto-PVFx and auto-AAC with incident fractures over a mean (SD) follow-up of 3.7 (2.2) years, adjusted for each other and other risk factors. At baseline, 17% (n = 1881) had auto-PVFx and 27% (n = 2974) had a high level of auto-AAC (≥ 6 on scale of 0 to 24). Multivariable-adjusted hazard ratios (HR) for incident major osteoporotic fracture (95% CI) were 1.85 (1.59, 2.15) for those with compared with those without auto-PVFx, and 1.36 (1.14, 1.62) for those with high compared with low auto-AAC. The multivariable-adjusted HRs for incident hip fracture were 1.62 (95% CI, 1.26 to 2.07) for those with compared to those without auto-PVFx, and 1.55 (95% CI, 1.15 to 2.09) for those high auto-AAC compared with low auto-AAC. The 5-year cumulative incidence of major osteoporotic fracture was 7.1% in those with no auto-PVFx and low auto-AAC, 10.1% in those with no auto-PVFx and high auto-AAC, 13.4% in those with auto-PVFx and low auto-AAC, and 18.0% in those with auto-PVFx and high auto-AAC. While physician manual review of images in clinical practice will still be needed to confirm image quality and provide clinical context for interpretation, simultaneous automated ascertainment of auto-PVFx and auto-AAC can aid fracture risk assessment.

Individuals with calcification of their abdominal aorta (AAC) and vertebral fractures seen on lateral spine bone density images (easily obtained as part of a bone density test) are much more likely to have subsequent fractures. Prior studies have not shown if both AAC and prior vertebral fracture both contribute to fracture prediction in routine clinical practice. Additionally, a barrier to using these images to aid fracture risk assessment at the time of bone density testing has been the need for expert readers to be able to accurately detect both AAC and vertebral fractures. We have developed automated computer methods (using artificial intelligence) to accurately detect vertebral fracture (auto-PVFx) and auto-AAC on lateral spine bone density images for 11 013 older individuals having a bone density test in routine clinical practice. Over a 5-year follow-up period, 7.1% of those with no auto-PVFx and low auto-AAC, 10.1% of those with no auto-PVFx and high auto-AAC, 13.4% of those with auto-PVFx and low auto-AAC, and 18.0% of those with auto-PVFx and high auto-AAC had a major osteoporotic fracture. Auto-PVFx and auto-AAC, ascertained simultaneously on lateral spine bone density images, both contribute to the risk of subsequent major osteoporotic fractures in routine clinical practice settings.

Machine learning for abdominal aortic calcification assessment from bone density machine-derived lateral spine images.

BACKGROUND: Lateral spine images for vertebral fracture assessment can be easily obtained on modern bone density machines. Abdominal aortic calcification (AAC) can be scored on these images by trained imaging specialists to assess cardiovascular disease risk. However, this process is laborious and requires careful training. METHODS: Training and testing of model performance of the convolutional neural network (CNN) algorithm for automated AAC-24 scoring utilised 5012 lateral spine images (2 manufacturers, 4 models of bone density machines), with trained imaging specialist AAC scores. Validation occurred in a registry-based cohort study of 8565 older men and women with images captured as part of routine clinical practice for fracture risk assessment. Cox proportional hazards models were used to estimate the association between machine-learning AAC (ML-AAC-24) scores with future incident Major Adverse Cardiovascular Events (MACE) that including death, hospitalised acute myocardial infarction or ischemic cerebrovascular disease ascertained from linked healthcare data. FINDINGS: The average intraclass correlation coefficient between imaging specialist and ML-AAC-24 scores for 5012 images was 0.84 (95% CI 0.83, 0.84) with classification accuracy of 80% for established AAC groups. During a mean follow-up 4 years in the registry-based cohort, MACE outcomes were reported in 1177 people (13.7%). With increasing ML-AAC-24 scores there was an increasing proportion of people with MACE (low 7.9%, moderate 14.5%, high 21.2%), as well as individual MACE components (all p-trend <0.001). After multivariable adjustment, moderate and high ML-AAC-24 groups remained significantly associated with MACE (HR 1.54, 95% CI 1.31-1.80 & HR 2.06, 95% CI 1.75-2.42, respectively), compared to those with low ML-AAC-24. INTERPRETATION: The ML-AAC-24 scores had substantial levels of agreement with trained imaging specialists, and was associated with a substantial gradient of risk for cardiovascular events in a real-world setting. This approach could be readily implemented into these clinical settings to improve identification of people at high CVD risk. FUNDING: The study was supported by a National Health and Medical Research Council of Australia Ideas grant and the Rady Innovation Fund, Rady Faculty of Health Sciences, University of Manitoba.

Development of a manufacturer-independent convolutional neural network for the automated identification of vertebral compression fractures in vertebral fracture assessment images using active learning.

BACKGROUND: Convolutional neural networks (CNNs) can identify vertebral compression fractures in GE vertebral fracture assessment (VFA) images with high balanced accuracy, but performance against Hologic VFAs is unknown. To obtain good classification performance, supervised machine learning requires balanced and labeled training data. Active learning is an iterative data annotation process with the ability to reduce the cost of labeling medical image data and reduce class imbalance. PURPOSE: To train CNNs to identify vertebral fractures in Hologic VFAs using an active learning approach, and evaluate the ability of CNNs to generalize to both Hologic and GE VFA images. METHODS: VFAs were obtained from the OsteoLaus Study (labeled Hologic Discovery A, n = 2726), the Manitoba Bone Mineral Density Program (labeled GE Prodigy and iDXA, n = 12,742), and the Canadian Longitudinal Study on Aging (CLSA, unlabeled Hologic Discovery A, n = 17,190). Unlabeled CLSA VFAs were split into five equal-sized partitions (n = 3438) and reviewed sequentially using active learning. Based on predicted fracture probability, 17.6% (n = 3032) of the unlabeled VFAs were selected for expert review using the modified algorithm-based qualitative (mABQ) method. CNNs were simultaneously trained on Hologic, GE dual-energy and GE single-energy VFAs. Two ensemble CNNs were constructed using the maximum and mean predicted probability from six separately trained CNNs that differed due to stochastic variation. CNNs were evaluated against the OsteoLaus validation set (n = 408) during the active learning process; ensemble performance was measured against the OsteoLaus test set (n = 819). RESULTS: The baseline CNN, prior to active learning, achieved 55.0% sensitivity, 97.9% specificity, 57.9% positive predictive value (PPV), F1-score 56.4%. Through active learning, 2942 CLSA Hologic VFAs (492 fractures) were added to the training data-increasing the proportion of Hologic VFAs with fractures from 4.2% to 12.5%. With active learning, CNN performance improved to 80.0% sensitivity, 99.7% specificity, 94.1% PPV, F1-score 86.5%. The CNN maximum ensemble achieved 91.9% sensitivity (100% for grade 3 and 95.5% for grade 2 fractures), 99.0% specificity, 81.0% PPV, F1-score 86.1%. CONCLUSION: Simultaneously training on a composite dataset consisting of both Hologic and GE VFAs allowed for the development of a single manufacturer-independent CNN that generalized to both scanner types with good classification performance. Active learning can reduce class imbalance and produce an effective medical image classifier while only labeling a subset of available unlabeled image data-thereby reducing the time and cost required to train a machine learning model.

Feasibility of a generalized convolutional neural network for automated identification of vertebral compression fractures: The Manitoba Bone Mineral Density Registry.

BACKGROUND: Vertebral fracture assessment (VFA) images are acquired in dual-energy (DE) or single-energy (SE) scan modes. Automated identification of vertebral compression fractures, from VFA images acquired using GE Healthcare scanners in DE mode, has achieved high accuracy through the use of convolutional neural networks (CNNs). Due to differences between DE and SE images, it is uncertain whether CNNs trained on one scan mode will generalize to the other. PURPOSE: To evaluate the ability of CNNs to generalize between GE DE and GE SE VFA scan modes. METHODS: 12,742 GE VFA images from the Manitoba Bone Mineral Density Program, obtained between 2010 and 2017, were exported in both DE and SE modes. VFAs were classified by imaging specialists as fracture present or absent using the modified algorithm-based qualitative (mABQ) method. VFA scans were randomly divided into independent training (60%), validation (10%), and test (30%) sets. Three CNN models were constructed by training separately on DE only, SE only, and a composite dataset comprised of both SE and DE VFAs. All three trained CNN models were separately evaluated against both SE and DE test datasets. RESULTS: Good performance was seen for CNNs trained and evaluated on the same scan mode. DE scans used for both training and evaluation (DE/DE) achieved 87.9% sensitivity, 87.4% specificity, and an area under the receiver operating characteristic curve (AUC) of 0.94. SE scans used for both training and evaluation (SE/SE) achieved 78.6% sensitivity, 90.6% specificity, AUC = 0.92. Conversely, CNNs performed poorly when evaluated on scan modes that differed from their training sets (AUC = 0.58). However, a composite CNN trained simultaneously on both SE and DE VFAs gave performance comparable to DE/DE (82.4% sensitivity, 94.3% specificity, AUC = 0.95); and provided improved performance over SE/SE (82.2% sensitivity, 92.3% specificity, AUC = 0.94). Positive predictive value was higher with the composite CNN compared with models trained solely on DE (74.5% vs. 58.7%) or SE VFAs (68.6% vs. 62.9%). CONCLUSION: CNNs for vertebral fracture identification are highly sensitive to scan mode. Training CNNs on a composite dataset, comprised of both GE DE and GE SE VFAs, allows CNNs to generalize to both scan modes and may facilitate the development of manufacturer-independent machine learning models for vertebral fracture detection.

Machine learning for automated abdominal aortic calcification scoring of DXA vertebral fracture assessment images: A pilot study.

BACKGROUND: Abdominal aortic calcification (AAC) identified on dual-energy x-ray absorptiometry (DXA) vertebral fracture assessment (VFA) lateral spine images is predictive of cardiovascular outcomes, but is time-consuming to perform manually. Whether this procedure can be automated using convolutional neural networks (CNNs), a class of machine learning algorithms used for image processing, has not been widely investigated. METHODS: Using the Province of Manitoba Bone Density Program DXA database, we selected a random sample of 1100 VFA images from individuals qualifying for VFA as part of their osteoporosis assessment. For each scan, AAC was manually scored using the 24-point semi-quantitative scale and categorized as low (score < 2), moderate (score 2 to <6), or high (score ≥ 6). An ensemble consisting of two CNNs was developed, by training and evaluating separately on single-energy and dual-energy images. AAC prediction was performed using the mean AAC score of the two models. RESULTS: Mean (SD) age of the cohort was 75.5 (6.7) years, 95.5% were female. Training (N = 770, 70%), validation (N = 110, 10%) and test sets (N = 220, 20%) were well-balanced with respect to baseline characteristics and AAC scores. For the test set, the Pearson correlation between the CNN-predicted and human-labelled scores was 0.93 with intraclass correlation coefficient for absolute agreement 0.91 (95% CI 0.89-0.93). Kappa for AAC category agreement (prevalence- and bias-adjusted, ordinal scale) was 0.71 (95% CI 0.65-0.78). There was complete separation of the low and high categories, without any low AAC score scans predicted to be high and vice versa. CONCLUSIONS: CNNs are capable of detecting AAC in VFA images, with high correlation between the human and predicted scores. These preliminary results suggest CNNs are a promising method for automatically detecting and quantifying AAC.

Identification of Vertebral Fractures by Convolutional Neural Networks to Predict Nonvertebral and Hip Fractures: A Registry-based Cohort Study of Dual X-ray Absorptiometry.

Background Detection of vertebral fractures (VFs) aids in management of osteoporosis and targeting of fracture prevention therapies. Purpose To determine whether convolutional neural networks (CNNs) can be trained to identify VFs at VF assessment (VFA) performed with dual-energy x-ray absorptiometry and if VFs identified by CNNs confer a similar prognosis compared with the expert reader reference standard. Materials and Methods In this retrospective study, 12 742 routine clinical VFA images obtained from February 2010 to December 2017 and reported as VF present or absent were used for CNN training and testing. All reporting physicians were diagnostic imaging specialists with at least 10 years of experience. Randomly selected training and validation sets were used to produce a CNN ensemble that calculates VF probability. A test set (30%; 3822 images) was used to assess CNN agreement with the human expert reader reference standard and CNN prediction of incident non-VFs. Statistical analyses included area under the receiver operating characteristic curve, two-tailed Student t tests, prevalence- and bias-adjusted κ value, Kaplan-Meier curves, and Cox proportional hazard models. Results This study included 12 742 patients (mean age, 76 years ± 7; 12 013 women). The CNN ensemble demonstrated an area under the receiver operating characteristic curve of 0.94 (95% confidence interval [CI]: 0.93, 0.95) for VF detection that corresponded to sensitivity of 87.4% (534 of 611), specificity of 88.4% (2838 of 3211), and prevalence- and bias-adjusted κ value of 0.77. On the basis of incident fracture data available for 2813 patients (mean follow up, 3.7 years), hazard ratios adjusted for baseline fracture probability were 1.7 (95% CI: 1.3, 2.2) for CNN versus 1.8 (95% CI: 1.3, 2.3) for expert reader-detected VFs for incident non-VF and 2.3 (95% CI: 1.5, 3.5) versus 2.4 (95% CI: 1.5, 3.7) for incident hip fracture. Conclusion Convolutional neural networks can identify vertebral fractures on vertebral fracture assessment images with high accuracy, and these convolutional neural network-identified vertebral fractures predict clinical fracture outcomes. © RSNA, 2019 Online supplemental material is available for this article.