Computer-Aided Diagnosis for Lung Lesion in Companion Animals from X-ray Images Using Deep Learning Techniques

Abstract

X-ray radiography in animals has the difficulty of interpretation due to a variety of animals. This leads to image misinterpretation for a non-specialist veterinarian in some clinics that has no radiologist. Based on statistics of veterinary specialists in the US in 2018, the role of radiologist currently faces a shortage problem, especially in the fields of veterinary, which has only 4.2% from all of the other veterinarians. In this paper, we proposed an animal X-ray diagnosis application, namely Pet-X, focusing on the lung lesion problem which has difficulty in interpreting and need to be inspected in many respiratory and cardiovascular related cases. Pet-X automatically learns the sets of dogs and cats thoracic radiograph images, consisting of two positions which are in lateral and ventrodorsal position, pre-processes the images and generates the lung lesion diagnosis model using deep learning techniques (i.e., Convolutional neural networks). The diagnosis model is used to detect the possibility of abnormal lungs, and classify the abnormality in to any three lesion types of abnormal lungs (i.e., Alveolar, Interstitial and Bronchial). The proposed model could achieve a sensitivity 76%, specificity 83.3%, and accuracy 79.6% for lung lesion detection, and a sensitivity 81%, specificity 63.67%, and accuracy 72.3% for abnormal lung classification. Moreover, our application applied the class activation mapping technique to locate the abnormal regions in the images. Finally, Pet-X could assist the veterinarian and radiologist users to diagnose lung lesion in companion animals from X-ray images.