Classification of dog skin diseases using deep learning with images captured from multispectral imaging device

Background Dog-associated infections are related to more than 70 human diseases. Given that the health diagnosis of a dog requires expertise of the veterinarian, an artificial intelligence model for detecting dog diseases could significantly reduce time and cost required for a diagnosis and efficiently maintain animal health. Objective We collected normal and multispectral images to develop classification model of each three dog skin diseases (bacterial dermatosis, fungal infection, and hypersensitivity allergic dermatosis). The single models (normal image- and multispectral image-based) and consensus models were developed used to four CNN model architecture (InceptionNet, ResNet, DenseNet, MobileNet) and select well-performed model. Results For single models, such as normal image- or multispectral image-based model, the best accuracies and Matthew's correlation coefficients (MCCs) for validation data set were 0.80 and 0.64 for bacterial dermatosis, 0.70 and 0.36 for fungal infection, and 0.82 and 0.47 for hypersensitivity allergic dermatosis. For the consensus models, the best accuracies and MCCs for the validation set were 0.89 and 0.76 for the bacterial dermatosis data set, 0.87 and 0.63 for the fungal infection data set, and 0.87 and 0.63 for the hypersensitivity allergic dermatosis data set, respectively, which supported that the consensus models of each disease were more balanced and well-performed. Conclusions We developed consensus models for each skin disease for dogs by combining each best model developed with the normal and multispectral images, respectively. Since the normal images could be used to determine areas suspected of lesion of skin disease and additionally the multispectral images could help confirming skin redness of the area, the models achieved higher prediction accuracy with balanced performance between sensitivity and specificity.