C16 - Diagnostic Validation of a Vertebral Heart Score Machine Learning Algorithm for Canine Lateral Thoracic Radiographs

Abstract: Background: The vertebral heart score (VHS) is a measurement used to index the heart size to the thoracic vertebra and is a useful tool for identifying the presence of heart disease, staging heart disease, and providing prognostic information. There is increasing development of computer-aided algorithms to perform specific tasks, which can reduce radiograph reading times and help minimize human error.
Hypothesis/Objectives: The purpose of this study was to validate the use of a VHS algorithm compared to manual VHS scoring by three board-certified veterinary cardiologists.
Animals: A total of 1,200 canine lateral radiographs were randomly selected from a commercial telemedicine databank.
Methods: A convolutional neural network centered around semantic segmentation of relevant anatomical features was developed to predict heart size and vertebral bodies. These predictions were used to calculate the VHS score. Three board-certified cardiologists manually scored 400 images each using the traditional Buchanan method. Post-scoring, the cardiologists evaluated the algorithm for appropriate algorithm landmarks and overall image quality.
Results: The 95th percentile absolute error between the cardiologist VHS score and the algorithm VHS score was 1.05 vertebrae (95% CI: 0.97 – 1.20 vertebrae) with a mean error of -0.09 vertebrae (95% CI: -0.12 – -0.05 vertebrae) and an absolute mean error of 0.3 vertebrae. The model was well-calibrated across the predictive range.
Conclusions and Clinical Importance: We found the performance of the VHS algorithm comparable to that of three board certified cardiologists.