Performance of an Artificial Intelligence Convolution Neural Network Software for the Detection of Confirmed Heart Failure in Dogs and Cats

ABSTRACT An independent dataset was used to retrospectively test the ability of an artificial intelligence (AI) program to detect radiographic signs of right or left heart failure (HF) and combine these findings to make a radiographic assessment of HF in dogs and cats. Experimental groups included 105 confirmed cases of heart disease in failure and a control group of 40 confirmed cases of heart disease not in failure. The program had an 87.04% sensitivity and an 80% specificity for radiographically concluding HF in dogs. Compared to true findings, the program had a 98% sensitivity in detecting an interstitial pattern, 100% sensitivity in detecting cardiomegaly, and 95% sensitivity in detecting a vascular pattern among other findings. In cats, the program had a 96% sensitivity in detecting an interstitial pattern, 96% sensitivity in detecting cardiomegaly, and 94% sensitivity in detecting a vascular pattern, among other findings. Despite this high sensitivity performance, the program only combines these findings to conclude HF with a sensitivity of 9.8% and specificity of 95%. The AI software had a similar success rate compared to radiologists in identifying the individual findings of HF in both species. This software has reached a level of sophistication to identify the different abnormalities as accurately as a large radiology group and also combine these findings and arrive at the radiographic conclusion of HF in dogs. The performance of this program in cats highlights the importance for veterinarians to utilize their medical knowledge to interpret the AI report like any other ancillary diagnostic test.