Diagnostics of Canine Elbow Dysplasia using Deep Learning with Explainability Analysis

Elbow Dysplasia is a term used to describe the presence of one or several abnormalities involving the elbow joint. It is genetically inheritable and could in the worst case lead to lameness in dogs. In Norway, dogs can be screened of elbow dysplasia when they are a minimum of 12 months old. This thesis builds on the NMBU master thesis by Steiro, where Steiro used convolutional neural networks on dog elbow x-ray images to automatically diagnose elbow dysplasia. This thesis delved therefore deeper, to see if one could identify the severity grade of elbow dysplasia by experimenting with parameters in convolutional neural networks. There was a total of 7229 x-ray images collected from various clinics across Norway between 2018 and 2021, that were used for analysis in this thesis. EfficientNet models of different complexities and other parameters, such as type of loss function and learning rate were used for classification. There were four-class models which looked at one normal class and three classes of abnormal elbows with increasing severity, and three-class models which only looked at abnormal elbows, that were tested. The highest overall performing model in terms of the four-class models, had a test accuracy of 95.8% and a high test MCC of 0.805. On the other hand, the highest overall performing three-class model had a test accuracy of 76.4% combined with a test MCC of 0.643. In addition to experimenting with different loss functions and learning rates, two distinct pre-processing methods were tested to boost performance for the three-class models. The first technique was using images from three channels, where two of the channels had augmented versions of the original image. The second technique was binarization of the image dataset, where two of the three classes of abnormal elbows were merged, making it a binary problem. Lastly, explainability analysis was implemented on the highest overall performing three-class model, to assess if the model could have potential to be used in a clinical setting. This was done with a method called Variance of the Gradients, to understand which regions of the elbow joint most affected the model's predictions. This method proved that the model was not reliable, because the model often looked outside of the elbow joint, which is outside the intended region of interest.