Oral Presentation ANZBMS-MEPSA-ANZORS 2022

Identification of asymptomatic vertebral fracture using Artificial Intelligence (AI) (#20)

Huy G Nguyen 1 2 , Hoa T Nguyen 3 , Linh T.T Nguyen 4 , Thach S Tran 5 , Sai H Ling 1 , Lan T Ho-Pham 2 , Tuan V Nguyen 1 6
  1. School of Biomedical Engineering, The University of Technology Sydney, Sydney, NSW, Australia
  2. Bone and Muscle Research Group, Ton Duc Thang University, Ho Chi Minh, Vietnam
  3. Can Tho University of Medicine, Can Tho, Vietnam
  4. The 108 Military Central Hospital, Ha Noi, Vietnam
  5. Biology of Bone, The Garvan Institute of medical research, Sydney, NSW, Australia
  6. School of Population Health, University of New South Wales, Sydney, NSW, Australia

Vertebral fracture (VF) is common in the elderly and is associated with increased risks of subsequent fracture and mortality. More than two-thirds of VF are asymptomatic and thus not treated. Existing methods were hard-to-interpret or black-boxes and vertebral morphometric properties were not explored. Therefore, we sought to develop an artificial intelligent (AI) method for identifying asymptomatic VF.

We analysed 1,016 lateral spinal X-rays from the Vietnamese Osteoporosis Study. VF was diagnosed by a rheumatologist using the Genant semiquantitative method. Our AI algorithm extracts the edge of a vertebral body into six identifiable vertices used to quantify the loss in height (Branched-Unet). A vertebra is considered normal, mild, moderate or severe if the height loss was <20%, 20-24%, 25-40% and >40%, respectively. We evaluated the prognostic performance of the AI algorithm on vertebra-level and person-level. We calculated sensitivity, specificity, and area under the receiver (AUC) to assess the classification performance of our AI algorithm.

We trained Branched-Unet on 98 films, and validated on 55 films, and achieved a dice-coefficient (DSC) of 90. The prevalence of vertebral fracture ascertained by clinician was 3.2% (28 / 863). At the individual level (n=863 films), the algorithm achieved a sensitivity of 89% and specificity of 62%. At the vertebral level (n=5913 vertebrae), the algorithm yielded a sensitivity of 71% and specificity of 91%. Further analysis showed a good concordance for non-fracture (91%, 5364/5875), mild (82%, 19/23) and severe (60%, 3/5) but not for moderate fracture (50%, 5/10). The AUC was 85% for both levels.

These results suggest that our AI algorithm can accurately identify asymptomatic vertebral fractures on plain radiographs without sacrificing interpretability. The rapid and automated approach can improve the efficiency of vertebral fractures in high-volume setting.

Table 1. Concordance between clinician diagnosis and AI classification over 863 films or 5913 vertebrae

6299ac487fa96-Table+1.png