Impact of level of expertise versus the statistical tool on vertebral fracture assessment (VFA) readings in cohort studies

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.

The fracture predictive ability of lumbar spine BMD and TBS as calculated based on different combinations of the lumbar spine vertebrae

Lumbar spine bone mineral density (BMD) and trabecular bone score (TBS) are both calculated on L1-L4 vertebrae. This study investigated the ability to predict osteoporotic fractures of BMD and TBS as calculated based on all possible adjacent L1-L4 vertebrae combinations. Present findings indicate that L1-L3 is an optimal combination to calculate LS-BMD or TBS.

INTRODUCTION

Lumbar spine (LS) BMD and TBS are both assessed in the LS DXA scans in the same region of interest, L1-L4. We aimed to investigate the ability to predict osteoporotic fractures of all the possible adjacent LS vertebrae combinations used to calculate BMD and TBS and to evaluate if any of these combinations performs better at osteoporotic fracture prediction than the traditional L1-L4 combination.

METHODS

This study was embedded in OsteoLaus-women cohort in Switzerland. LS-DXA scans were performed using Discovery A System (Hologic). The incident vertebral fractures (VFs) and major osteoporotic fractures (MOFs) were assessed from VF assessments using Genant's method or questionnaires (non-VF MOF). We ran logistic models using TBS and BMD to predict MOF, VF, and non-VF MOF, combining different adjustment factors (age, fracture level, or BMD).

RESULTS

One thousand six hundred thirty-two women (mean ± SD) 64.4 ± 7.5 years, BMI 25.9 ± 4.5 kg/m², were followed for 4.4 years and 133 experienced MOF. The association of one SD decrease L1-L3 BMD with the odds ratios (ORs) of MOF was OR 1.32 (95%CI 1.15-1.53), L2-L4 BMD was 1.25 (95%CI 1.09-1.42), and L1-L4 BMD was 1.30 (95%CI 1.14-1.48). One SD decrease in L1-L3 TBS was more strongly associated with the odds of having a MOF (OR 1.64, 95% CI 1.34-2.00), than one SD decrease in L2-L4 TBS (OR 1.48, 95% CI 1.21-1.81), or in L1-L4 TBS (OR 1.60, CI 95% 1.32-1.95).

CONCLUSION

Current findings indicate that L1-L3 is an optimal combination for the TBS or LS-BMD calculation.

Assessment of Skeletal Strength: Bone Density Testing and Beyond

A bone fractures when a force applied to it exceeds its strength. Assessment of bone strength is an important component in determining the risk of fracture and guiding treatment decisions. Dual-energy X-ray absorptiometry is used to diagnosis osteoporosis, estimate fracture risk, and monitor changes in bone density. Fracture risk algorithms provide enhanced fracture risk predictability. Advanced technologies with computed tomography (CT) and MRI can measure parameters of bone microarchitecture. Mathematical modeling using CT data can evaluate the behavior of bone structures in response to external loading. Microindentation techniques directly measure the strength of outer bone cortex.

Vertebral Fracture Assessment in Postmenopausal Women With Postsurgical Hypoparathyroidism

CONTEXT

Hypoparathyroidism is a rare endocrine disorder whose skeletal features include suppression of bone turnover and greater volume and width of the trabecular compartment. Few and inconsistent data are available on the prevalence of vertebral fractures (VF).

OBJECTIVE

To evaluate the prevalence of VF assessed by vertebral fracture assessment (VFA) in postmenopausal women with chronic postsurgical hypoparathyroidism.

DESIGN

Cross-sectional study.

SETTING

Ambulatory referral center.

PATIENTS OR OTHER PARTICIPANTS

Fifty postmenopausal women (mean age 65.4 ± 9 years) with chronic postsurgical hypoparathyroidism and 40 age-matched healthy postmenopausal women (mean age 64.2 ± 8.6).

MAIN OUTCOME MEASURES

Lumbar spine, femoral neck, and total hip bone mineral density were measured by dual X-ray absorptiometry (Hologic Inc., USA) in all subjects. Site-matched spine trabecular bone score was calculated by TBS iNsight (Medimaps, Switzerland). Assessment of VF was made by VFA (iDXA, Lunar GE, USA) using the semiquantitative method and the algorithm-based qualitative assessment.

RESULTS

All-site BMD values were higher in the hypoparathyroid vs the control group. By VFA, we observed a 16% prevalence of VF in hypoparathyroid women vs 7.5% in control subjects. Among those with hypoparathyroidism who fractured, 5 (62.5%) had grade 1 wedge, 2 (25%) had grade 2 wedge, and 1 (12.5%) had grade 2 wedge and grade 2 biconcave VF. In the hypoparathyroid group, 57% with VFs and 32% without VFs had symptoms of hypoparathyroidism.

CONCLUSION

We demonstrate for the first time that in postmenopausal women with chronic postsurgical hypoparathyroidism, VFs are demonstrable by VFA despite normal BMD.

Vertebral fracture assessment (VFA) for monitoring vertebral reshaping in children and adolescents with osteogenesis imperfecta treated with intravenous neridronate

PURPOSE

The study was aimed at monitoring vertebral bodies changes with the use of Vertebral Fracture Assessment (VFA) in children and adolescents affected by osteogenesis imperfecta (OI) during treatment with intravenous neridronate.

METHODS

60 children and adolescents (35 males and 25 females; age 1-16 years) with OI type I, III and IV were included in the study. Intravenous neridronate was administered at the dose of 2 mg/kg every 3 months in all patients. Lumbar spine (LS) bone mineral density (BMD) and VFA by dual X-ray absorptiometry (DXA) were assessed every 6 months up to 24 months during treatment. VFA with vertebral morphometry (MXA) was used to calculate the three indices of vertebral deformity: wedging, concavity and crushing. Serum calcium, phosphate, parathyroid hormone (PTH), 25-hydroxy-vitamin D [25(OH)D], total alkaline phosphatase (ALP), bone alkaline phosphatase (BALP) and urinary C-terminal telopeptide of type 1 collagen (CTx) were measured at any time point.

RESULTS

Mean LS BMD values significantly increased at 24 months compared to baseline (p < 0.0001); the corresponding Z-score values were -1.28 ± 1.23 at 24 months vs -2.46 ± 1.25 at baseline; corresponding mean Bone Mineral Apparent Density (BMAD) values were 0.335 ± 0.206 vs 0.464 ± 0.216. Mean serum levels of ALP, BALP and CTx significantly decreased from baseline to 24 months. By MXA, we observed a significant 19.1% reduction of the mean wedging index of vertebral reshaping at 12 months, and 38.4% at 24 months (p < 0.0001) and of the mean concavity index (16.3% at 12 months and 35.9% at 24 months; p < 0.0001). Vertebral reshaping was achieved for 66/88 (75%) wedge fractures and 59/70 (84%) concave fractures, but there were 4 incident mild fractures. Finally, VF rate was reduced at 24 months compared to baseline: 37/710 (5.2%) vs 158/710 (22.2%).

CONCLUSION

Our study demonstrates the utility of VFA as a safe and alternative methodology in the follow-up of children and adolescents with OI.

Grade 1 Vertebral Fractures Identified by Densitometric Lateral Spine Imaging Predict Incident Major Osteoporotic Fracture Independently of Clinical Risk Factors and Bone Mineral Density in Older Women

Because prevalent vertebral fracture (VF) is a strong predictor of future fractures, they are important to identify in clinical practice as osteoporosis medications are effective and can be used to reduce fracture risk in postmenopausal women with VF. Lateral spine imaging (LSI) with dual-energy X-ray absorptiometry (DXA) can be used to diagnose VFs accurately but is not widespread in clinical practice. The prognostic value of grade 1 (20% to 25% compression) VFs diagnosed by LSI with DXA has been insufficiently studied. The aim of this study was to determine if grade 1 VF is associated with incident fracture in older women. Sahlgrenska University Hospital Prospective Evaluation of Risk of Bone Fractures (SUPERB) is a population-based study of 3028 older women from Gothenburg, Sweden. Included women were 75 to 80 years of age at baseline, answered questionnaires, and were scanned with DXA (Discovery A, Hologic, Waltham, MA, USA). LSI was used to diagnose VFs, which were classified using the Genant semiquantitative method. Cox regression models were used to estimate the association between VFs at baseline and X-ray-verified incident fractures, with adjustment for confounders. Women with a grade 1 VF (n = 264) or a grade 2-3 VF (n = 349) were compared with women without any fracture (n = 1482). During 3.6 years (median, interquartile range [IQR] 1.5 years) of follow-up, 260 women had any incident fracture and 213 a major osteoporotic fracture (MOF). Women with only grade 1 VF had increased risk of any fracture (hazard ratio [HR] = 1.67; 95% confidence interval [CI] 1.18-2.36) and MOF (HR = 1.86; 95% CI 1.28-2.72). For MOF, this association remained after adjustment for clinical risk factors and femoral neck bone mineral density (BMD). In conclusion, grade 1 VFs were associated with incident MOF, also after adjustment for clinical risk factors and BMD, indicating that all VF identified by DXA should be considered in the evaluation of fracture risk in older women. © 2020 The Authors. Journal of Bone and Mineral Research published by American Society for Bone and Mineral Research..

Corrigendum to how to define an osteoporotic vertebral fracture

[This corrects the article DOI: 10.21037/qims.2019.09.10.].

Clinical Performance of the Updated Trabecular Bone Score (TBS) Algorithm, Which Accounts for the Soft Tissue Thickness: The OsteoLaus Study

Regional soft tissue may have a noise effect on trabecular bone score (TBS) and eventually alter its estimate. The current TBS software (TBS iNsight®) is based on an algorithm accounting for body mass index (BMI) (TBS_v3.03 ). We aimed to explore the updated TBS algorithm that accounts for soft tissue thickness (TBS_v4.0 ). This study was embedded in the OsteoLaus cohort of women in Lausanne, Switzerland. Hip and lumbar spine (LS) dual-energy X-ray absorptiometry (DXA) scans were performed using Discovery A System (Hologic). The incident major osteoporotic fractures (MOFs) were assessed from vertebral fracture assessments using Genant's method (vertebral MOF) or questionnaires (nonvertebral MOF). We assessed the correlations of bone mineral density (BMD) or TBS with body composition parameters; MOF prediction ability of both versions of TBS; and the differences between Fracture Risk Assessment Tool (FRAX) adjusted for TBS_v3.03 or TBS_v4.0 . In total, 1362 women with mean ± SD age 64.4 ± 7.5 years and mean ± SD BMI 25.9 ± 4.5 kg/m² were followed for 4.4 years and 132 experienced an MOF. All the anthropometric measurements of our interest were positively correlated with LS, femoral neck, or hip BMD and TBS_v4.0 ; whereas with TBS_v3.03 their correlations were negative. In the models adjusted for age, soft tissue thickness, osteoporotic treatment, and LS-BMD, for each SD decline in TBS_v3.03 , there was a 43% (OR 1.43; 95% CI, 1.12 to 1.83) increase in the odds of having MOF; whereas for each SD decline in TBS_v4.0 , there was a 54% (OR 1.54; 95% CI, 1.18 to 2.00) increase in the odds of having an MOF. Both FRAXs were very strongly correlated and the mild differences were present in the already high-risk women for MOF. This study shows that TBS_v4.0 overcomes the debatable residual negative correlation of the current TBS with body size and composition parameters, postulating itself as free from the previously acknowledged technical limitation of TBS. © 2019 American Society for Bone and Mineral Research.

Vertebral Fracture Assessment Increases Use of Pharmacologic Therapy for Fracture Prevention in Clinical Practice

The impact of vertebral fracture assessment (VFA) on lateral spine images in clinical practice on subsequent patient use of fracture prevention medication is unknown. Our objective was to determine the association of prevalent vertebral fracture identified on bone density lateral spine images (positive VFA) with subsequent use of fracture prevention therapy in usual clinical practice, using the Manitoba Bone Density Program database prospective observational cohort. Since 2010, targeted VFA imaging has been done at the time of bone densitometry in Manitoba for 21% of women and men meeting criteria based on age, bone mineral density (BMD), height loss, and glucocorticoid use. Among 6652 treatment-naive individuals with at least 90 days follow-up who had VFA imaging, 923 (13.9%) had one or more definite vertebral fractures identified using a modified algorithm-based qualitative (ABQ) method. For those with a positive VFA, their bone density reports stated the patient was at high risk of subsequent fracture and qualified for fracture prevention therapy. Subsequent osteoporosis treatment initiated within the next 12 months was identified using population-based pharmacy data. Logistic regression models were used to estimate the association of positive VFA with subsequent prescription (Rx), compared to negative VFA. Fracture prevention medication was started by 2127 (32%) individuals, 52.3% with positive versus 28.4% with negative VFA (p value <0.001). This association was substantially stronger in those designated (before VFA results were known) to have low or moderate fracture risk compared to high fracture risk (interaction p value <0.001), and in those with osteopenia (OR 4.51; 95% CI, 3.48 to 5.85) compared to those with osteoporosis by BMD criteria (OR 1.72; 95% CI, 1.43 to 2.08, interaction p value <0.001). Targeted VFA imaging at the time of bone densitometry substantially improves identification of those at high fracture risk and fracture prevention medication use among those with prevalent vertebral fracture. © 2019 American Society for Bone and Mineral Research. © 2019 American Society for Bone and Mineral Research.

How to define an osteoporotic vertebral fracture?

A vertebral deformity (VD) is not always a vertebral fracture (VF). Because of lack of a completely satisfactory "gold standard", there is no consensus on the exact definition of a VF. Therefore, it may sometimes be difficult, especially in mild cases, to discriminate the prevalent VF from a non-fracture deformity or short vertebral height (SVH). A combined standardized approach based on qualitative and semiquantitative (SQ) vertebral assessment may be the most option to correctly identify a VD as a VF. However this visual approach for VF identification is subjective, therefore it is mandatory an adequate training and experience of radiologist to reach a good sensitivity and specificity. Vertebral morphometry, objective and reproducible method, could be used only to evaluate the severity of VFs but requires the availability of reference values of vertebral height ratios. There is actually an evidentiary basis for suggesting that a qualitative approach by expert radiologists to morphological vertebral assessment, combined SQ and morphometric methods seem to be the preferred option for the correct diagnosis of VF as endplate or/and cortex fracture (ECF) or severe vertebral height loss.

Prevalent vertebral fracture on bone density lateral spine (VFA) images in routine clinical practice predict incident fractures

PURPOSE

The predictive validity of vertebral fracture assessment (VFA) on bone density lateral spine images to identify prevalent vertebral fractures in routine clinical practice has not been established. Our objective was to estimate the associations of prevalent vertebral fracture identified on VFA images in routine practice with incident hip, all non-vertebral, major osteoporotic, and clinical vertebral fractures, using the Manitoba Bone Density database.

METHODS

From 2010 onward, 9972 men and women (mean age [SD] 76 [6.9] years) had VFA images obtained at the time of bone densitometry that were interpreted for vertebral fracture by the clinicians reading the bone density tests. Definite and possible prevalent vertebral fractures, respectively, were identified in 1575 (15.8%) and 293 (2.9%) using a modified Algorithm Based Qualitative method. We ascertained incident fractures using Manitoba provincial health databases over a mean 2.8 (SD 1.7) years and used Cox proportional hazards models to estimate the associations of prevalent vertebral fractures with incident fractures.

RESULTS

Compared to no prevalent vertebral fracture, those with definite prevalent vertebral fracture had higher hazard ratios for incident hip (HR 1.95, 95% C.I. 1.45 to 2.62), non-vertebral (HR 1.99, 95% C.I. 1.68 to 2.35), and clinical vertebral fracture (HR 2.68, 95% C.I. 1.69 to 4.23) adjusted for age, bone mineral density, body mass index, prior fracture, parental hip fracture, glucocorticoid use, alcohol use, smoking, and rheumatoid arthritis. These associations did not vary by FRAX fracture risk estimates or bone mineral density category.

CONCLUSION

Prevalent vertebral fractures identified on densitometric VFA images in routine clinical practice are strongly associated with incident fractures, and this study is the first to show this using any lateral spine imaging modality outside of research settings. These findings are strong evidence supporting the targeted use of densitometric VFA imaging among post-menopausal women and older men referred for bone densitometry.

Vertebral Fracture Identification as Part of a Comprehensive Risk Assessment in Patients with Osteoporosis

PURPOSE OF REVIEW

To review current evidence regarding the vertebral fracture prevalence, the accuracy of vertebral fracture identification on current imaging technologies, and the potential impact of vertebral fracture identification on fracture risk.

RECENT FINDINGS

Important new studies have clarified the features of prevalent vertebral fracture that most strongly predict incident fractures. Age- and sex-stratified estimates of vertebral fracture prevalence on densitometric lateral spine images in the US population are now available. The accuracy of densitometric vertebral fracture assessment, how computed tomography scans and other spinal images obtained for indications other than vertebral fracture assessment can be leveraged to detect prevalent vertebral fractures, and the potential impact of vertebral fracture assessment on patient and provider fracture risk management behavior have been clarified. Substantial progress has been made regarding screening strategies using lateral spine imaging to detect prevalent vertebral fracture in the older population. Further research regarding implementation of these strategies in clinical practice and their impact on clinical outcomes is needed.