Artificial intelligence (AI) has the capability to accurately and reliably detect thoracolumbar fractures on sagittal radiographs, according the findings from a retrospective study that was presented during a Best of Show session at the 2021 Eurospine annual meeting (6–8 October, Vienna, Austria) by Enrico Gallazzi (Centro Specialistico Ortopedico Traumatologico G. Pini, Milan, Italy).
Gallazzi and his fellow researchers note that the use of deep learning “could enhance the detection of fractures on simple radiographic projections, providing benefits to settings in which tomographic images are not readily available such as emergency condition or developing countries and assisting the clinician in identifying patients in need of second level imaging.
“Moreover, through the analysis of heatmaps, it was qualitatively verified that the model is in fact identifying the correct location of the fracture within the vertebral body,” they add.
The study examined whether or not deep learning methods could be used to reliably and accurately detect thoracolumbar fractures in sagittal radiographs of the spine.
Sagittal radiographs, CT and MRI scans of the thoracolumbar spine of 362 patients exhibiting traumatic vertebral fractures were collected. Three expert spine surgeons labelled the dataset by drawing a Region of Interest (ROI) around each fractured and non-fractured vertebra on the sagittal x-rays, indicating the corresponding class (fracture/no-fracture), and using CT and MRI to confirm the presence of the fracture.
From this dataset, 279 x-ray images of fractured vertebrae and 288 showing no fracture were annotated. The dataset was then used to train, validate and test two deep learning classifiers based on the ResNet18 and VGG architectures. Heatmaps indicating which parts of the images led the model to classify the vertebra as ‘fracture’ or ‘no fracture’ were obtained and then evaluated by the same surgeons.
Among the 52 images constituting the test set, accuracies of 88% and 84% were obtained with ResNet and VGG, respectively. Sensitivity was 89% with both architectures but ResNet had a significantly higher specificity of 88% compared to 79% of the VGG. A total of 42 of the 52 heatmaps (81%) were judged to correctly indicate the fracture location.
Speaking to Spinal News International, Gallazzi said: “This study showed how AI could be used to improve a weak point in the management of traumatic thoracolumbar vertebral fractures, given that the rate of missed fractures in plain x-rays as reported in the literature is up to 60%.
“Moreover, this approach could greatly benefit developing countries, given the low availability of second level imaging as CT scans. We aim to keep improving the performance of our algorithm and extending the analysis to fracture classification. As a group, we are sure of the potential of this techniques and are excited to continue this research.”