Oral Presentation 29th Australian and New Zealand Bone and Mineral Society Annual Scientific Meeting 2019

Utility of an automated vertebral fracture detection software (#22)

Nithin Kolanu 1 2 , Elizabeth Silverstone 1 , Bao Ho 1 , Hiep Pham 1 , Ash Hansen 1 , Emma Pauley 1 , Anna Quirk 1 , Sarah Sweeney 1 , Jackie Center 1 2 3 , Nicholas Pocock 1 2 3
  1. St Vincent's Hospital, Darlinghurst, NSW, Australia
  2. Garvan Institute of Medical Research, Sydney, NSW
  3. University of New South Wales, Sydney, NSW, Australia

Introduction:

An ageing population is resulting in increased prevalence of osteoporotic fractures[1] which are a risk factor for subsequent fractures[2,3]. Vertebral fractures are common, frequently asymptomatic and often present in Computed Tomography (CT) scans performed as part of clinical care for unrelated conditions. Reporting frequency of these vertebral fractures varies.

Machine learning algorithms can diagnose fractures from previous radiological images [4]. Fracture detection algorithms [5] have the potential to improve detection and reporting of asymptomatic vertebral fractures but further testing is essential [6] before widespread use.

Methods:

Abdominal and thoracic CT scans from a tertiary hospital, performed over 3 months, in subjects aged over 50 years were reviewed by a radiologist (n =2237) and by using image analysis software (Zebra Medical Vision®, n = 1991).  Sensitivity and specificity of the software, compared to the radiologist (current standard of care), were calculated for detecting any vertebral fracture (n =1626), and separately for the detection of Genant grade 2 and 3 fractures (n =1487). All potential false positives and false negatives were reviewed by a second imaging specialist and any discrepancies between specialist 1 and 2 were adjudicated by a 3rd imaging specialist to establish a final diagnosis.

Results:

Image analytics software had sensitivity of 53% and specificity of 92% for detection of any vertebral fracture when compared to radiologist review. Sensitivity and specificity for detection of Genant Grade 2 and 3 fractures was 64% and 92% respectively. Accuracy for any vertebral fracture detection, and for detection of Genant grade 2 or 3 fractures, was 82% and 87% respectively.

Conclusion:

Image analytics software has high specificity with moderate sensitivity, to detect vertebral fractures in CT scans performed for any indication. It potentially is a useful tool to assist clinicians and radiologists to improve detection and reporting of vertebral fractures.

  1. 1. Watts JJ, Abimanyi-Ochom J, Sanders KM, Osteoporosis costing all Australians A new burden of disease analysis – 2012 to 2022 (2013)
  2. 2. Kanis JA, Johnell O, De Laet C, Johansson H, Oden A, Delmas PD, Eisman JA, Fujiwara S, Garnero P,Kroger H, McCloskey EV, Mellstrom D, Melton LJ, Pols H, Reeve J, Silman A, Tenenhouse A. A metaanalysis of previous fracture and subsequent fracture risk. Bone 2004;35(2):375-82
  3. 3. Melton III LJ, Atkinson EJ, Cooper C, O'Fallon WM, Riggs BL. Vertebral fractures predict subsequent fractures. Osteoporosis Int 1999;10:214-21
  4. 4. Kim DH, MacKinnon T. Artificial intelligence in fracture detection: transfer learning from deep convolutional neural networks. Clinical Radiology 2018; 73(5);439-45
  5. 5. Massat MB. Artificial intelligence in radiology: Hype or hope?. Appl Radiol. 2018;47(3):22-26
  6. 6. Burns JE, Yao J, Munoz H, Summers RM. Automated detection, localization, and classification of Traumatic Vertebral Body Fractures in the Thoracic and lumbar spine at CT. Radiology 2016;278(1):64-73