Comparison of radiological interpretation made by veterinary radiologists and AI software for canine and feline radiographic studies.

Human-based radiologic interpretation suffers from both noise and bias. Artificial Intelligence (AI) has the potential to reduce bias and increase diagnostic availability. Purpose of this study was to analyze how commercial AI software dedicated to veterinary radiology compares to the performance of veterinary radiologists. Our hypotheses were that the mean diagnostic accuracy of AI will be higher than the mean diagnostic accuracy of veterinary radiologists and that the diagnostic accuracy of AI is higher than the diagnostic accuracy of any radiologist. Fifty canine and feline radiographic studies in DICOM format were anonymized and reported by 11 boardcertified veterinary radiologists and processed with commercial AI software dedicated to small animal radiography (SignalRAY, SignalPET Dallas, TX, USA). The diagnostic were recorded and coded. The mean sensitivities, specificities and accuracies were high for both veterinary radiologists (0.5787; 0.9650; 0.8733) and the AI software (0.7008; 0.9402; 0.9204).
The AI software outperformed the below-mean veterinary radiologists (p<0.0002), confirming our first hypothesis. There was no significant difference in diagnostic accuracy between the AI software and the best-performing veterinary radiologist (p=0.0051), rejecting our second hypothesis. AI performed better in low-noise settings and exhibited different strengths in lownoise settings and high-noise settings. Given the unique strengths of human experts and AI, as well as the differences in sensitivity versus specificity and low-noise versus high-noise settings, AI is likely to best complement rather than substitute human experts. Hence, it would be worth exploring how AI could be provided to human experts to enhance their performance.