Medical image analysis based on deep learning is a rapidly advancing field in veterinary diagnostics. The aim of this retrospective diagnostic accuracy study was to develop and assess a convolutional neural network (CNN, EfficientNet) to evaluate elbow radiographs from dogs screened for elbow dysplasia.
An auto-cropping tool based on the deep learning model RetinaNet was developed for radiograph preprocessing to crop the radiographs to the region of interest around the elbow joint. A total of 7229 radiographs with corresponding International Elbow Working Group scoring were included for training (n = 4000), validation (n = 1000), and testing (n = 2229) of CNN models for elbow diagnostics. The radiographs were classified in a binary manner as normal (negative class) or abnormal (positive class), where abnormal radiographs had various severities of osteoarthrosis and/or visible primary elbow dysplasia lesions.
Explainable artificial intelligence analysis were performed on both correctly and incorrectly classified radiographs using VarGrad heatmaps to visualize regions of importance for the CNN model's predictions. The highest-performing CNN model showed excellent test accuracy, sensitivity, and specificity, all achieving a value of 0.98. Explainability analysis showed frequent highlighting along the margins of the anconeal process of both correctly and incorrectly classified radiographs. Uncertainty estimation using entropy to characterize the uncertainty of the model predictions showed that radiographs with ambiguous predictions could be flagged for human evaluation.
Our study demonstrates robust performance of CNNs for detecting abnormal elbow joints in dogs screened for elbow dysplasia.
