Background/Purpose: Vertebral Compression Fractures are common in patients above age 50, but are often undiagnosed.  Patients with one VCF are at higher risk of other osteoporotic fractures. Zebra Medical Imaging developed a VCF detection algorithm, utilizing a combination of traditional machine vision segmentation and convolutional neural network (CNN) technology,  to detect VCFs from evaluating  CT images of the chest, and/or abdomen/pelvis.

Methods: We conducted an independent and blinded validation study to estimate the operating characteristics of the Zebra VCF detection algorithm in identifying VCFs on de-identified data from previously completed CT scans of chest and/or abdomen/pelvis from 1200 women and men aged 50 or older who had those scans (for multiple reasons) at the clinics and hospitals affiliated with the Cedars Sinai Medical Center. Each set of scans were independently read by two of three board certified, practicing neuroradiologists to identify and grade VCF at each evaluable vertebra (using the semiquantitative scale of  Genant and colleagues).  When there was disagreement between radiologists, the respective scans were reviewed by a senior neuroradiologist who provided a final evaluation.  The final determination of presence and severity of VCF by the neuroradiologists was used as the reference standard. The Zebra VCF detection algorithm evaluated the CT scans in a separate workstream from that used by the neuroradiologists (The algorithm and neuroradiologists were blind to each other’s evaluations).

Results: The Zebra VCF algorithm was not able to evaluate CT scans for 113 patients.  Of the remaining 1087 CT patients, 588 (54%) were women.  Median age was 73 (range 51, 102; interquartile range 66, 81).  The four neuroradiologists who evaluated the CT scans each had over 10 years of experience in neuroradiology. For the 1087 Zebra evaluated patients, 227 had at least one VCF (90 with mild VCF, 81 with moderate VCF, and 56 with severe VCF; 115 of the 1087 Zebra evaluated patients (10.6%) presented with two or more VCFs).  The sensitivity and specificity of the Zebra VCF algorithm in diagnosing any VCF were 0.66 (95% confidence interval 0.59, 0.72) and 0.90 (95% confidence interval 0.88, 0.92) respectively; and for diagnosing moderate/severe VCF were 0.78 (95% confidence interval  0.70, 0.85) and 0.87 (95% confidence interval 0.85, 0.89) respectively.

Conclusion: The Zebra VCF algorithm works to identify approximately three-quarters of moderate to severe VCF in patients, aged 50 and above, who receive CT scans for other reasons. Implementing the Zebra VCF algorithm within radiology systems may help to identify patients at increased fracture risk and could support the diagnosis of osteoporosis, and thus be a valuable adjunct for population health.

Comparison of diagnoses by radiologists and Zebra VCF detection algorithm (Operating Characteristics)

« Back to ACR Convergence 2020

ACR Meeting Abstracts - https://acrabstracts.org/abstract/validation-of-a-deep-learning-based-algorithm-to-diagnose-vertebral-compression-fractures/