Investigation of the Accuracy and Contributing Factors of AI-Based Diagnosis of Urothelial Carcinoma in Canine Abdominal Radiography.

Abstract

Urothelial carcinoma (UC) is a highly malignant urinary cancer of the transitional epithelium in dogs. Recent advances in artificial intelligence (AI) and machine learning have shown substantial potential in veterinary medicine. The purpose of this study was to evaluate the accuracy of AI-based software in detecting UC in dogs using abdominal radiography and to identify factors that influence the sensitivity of AI-based diagnosis. Dogs underwent abdominal radiography, and ultrasound was retrospectively retrieved. Dogs with histologically confirmed UC and ultrasound changes were included in UC training and UC validation groups, whereas dogs without clinical suspicion of urinary neoplasia and without ultrasound findings consistent with UC were included as non-UC training and non-UC validation groups. Histological and imaging findings of UC were recorded. A convolutional neural network (CNN) was trained with 500 studies from the UC training and 500 studies from the non-UC training groups. For validation, an additional 185 studies from the UC validation and 180 studies from the non-UC validation groups were used to provide AI-based diagnosis of UC by the trained CNN. The sensitivity, specificity, and accuracy of the AI-based diagnosis of UC were 69%, 67%, and 68%, respectively. The software showed higher sensitivity in detecting more severe UC with mineralization. However, medial iliac lymphadenomegaly and ureteral obstruction did not improve the sensitivity of AI-based diagnosis. In conclusion, well-trained CNN demonstrated moderate accuracy for detecting UC using abdominal radiographs, with higher sensitivity in cases with more advanced disease. The unexpectedly superior performance of the ventrodorsal (VD) view warrants further investigation.