Guillermo F Sanchez Rosenberg, Andrea Cina, Giuseppe Rosario Schiro, Pietro Domenico Giorgi, Boyko Gueorguiev, Mauro Alini, Peter Varga, Fabio Galbusera, Enrico Gallazzi
2021 medRxiv   pre-print
context Traumatic thoracolumbar (TL) fractures are frequently encountered in emergency rooms. Sagittal and anteroposterior radiographs are the first step in the trauma routine imaging. Up to 30% of TL fractures are missed in this imaging modality, thus requiring a CT and/or MRI to confirm the diagnosis. A delay in treatment leads to increased morbidity, mortality, exposure to ionizing radiation and financial burden. Fracture detection with Machine Learning models has achieved expert level
more » ... mance in previous studies. Reliably detecting vertebral fractures in simple radiographic projections would have a significant clinical and financial impact. Purpose To develop a deep learning model that detects traumatic fractures on sagittal radiographs of the TL spine. Study design/setting Retrospective Cohort study. Methods We collected sagittal radiographs, CT and MRI scans of the TL spine of 362 patients exhibiting traumatic vertebral fractures. Cases were excluded when CT and/or MRI where not available. The reference standard was set by an expert group of three spine surgeons who conjointly annotated the sagittal radiographs of 171 cases. CT and/or MRI were reviewed to confirm the presence and type of the fracture in all cases. 302 cropped vertebral images were labelled -fracture- and 328 -no fracture-. After augmentation, this dataset was then used to train, validate, and test deep learning classifiers based on ResNet18 and VGG16 architectures. To ensure that the prediction of the model was based on the correct identification of the fracture zone, an Activation Map analysis was conducted. Results Vertebras T12 to L2 were the most frequently involved, accounting for 48% of the fractures. A4, A3 and A1 were the most frequent AO Spine fracture types. Accuracies of 88% and 84% were obtained with ResNet18 and VGG16 respectively. The sensitivity was 89% with both architectures but ResNet18 showed a higher specificity (88%) compared to VGG16 (79%). The fracture zone was precisely identified in 81% of the heatmaps. Conclusions Our AI model can accurately identify anomalies suggestive of vertebral fractures in sagittal radiographs by precisely identifying the fracture zone within the vertebral body. Clinical significance Clinical implementation of a diagnosis aid tool specifically trained for TL fracture identification is anticipated to reduce the rate of missed vertebral fractures in emergency rooms.
doi:10.1101/2021.05.09.21256762 fatcat:3nu2qfjjyjdjfoj7yvpzv3j3hm