Combining ultrasound radiomics, complete blood count, and serum biochemical biomarkers for diagnosing intestinal disorders in cats using machine learning.

This retrospective analytical observational cohort study aimed to model and predict the classification of feline intestinal diseases from segmentations of a transverse section from small intestine ultrasound (US) image, complete blood count (CBC), and serum biochemical profile data using a variety of machine-learning approaches. In 149 cats from three institutions, images were obtained from cats with biopsy-confirmed small cell epitheliotropic lymphoma (lymphoma), inflammatory bowel disease (IBD), no pathology ("healthy"), and other conditions (warrant a biopsy for further diagnosis). CBC, blood serum chemistry, small intestinal ultrasound, and small intestinal biopsy were obtained within a 2-week interval. CBC and serum biomarkers and radiomic features were combined for modeling. Four classification schemes were investigated: (1) normal versus abnormal; (2) warranting or not warranting a biopsy; (3) lymphoma, IBD, healthy, or other conditions; and (4) lymphoma, IBD, or other conditions. Two feature selection methods were used to identify the top 3, 5, 10, and 20 features, and six machine learning models were trained. The average (95% CI) performance of models for all combinations of features, numbers of features, and types of classifiers was 0.886 (0.871-0.912) for Model 1 (normal vs. abnormal), 0.751 (0.735-0.818) for Model 2 (biopsy vs. no biopsy), 0.504 (0.450-0.556) for Model 3 (lymphoma, IBD, healthy, or other), and 0.531 (0.426-0.589), for Model 4 (lymphoma, IBD, or other). Our findings suggest model accuracies above 0.85 can be achieved in Model 1 and 2, and that including CBC and biochemistry data with US radiomics data did not significantly improve accuracy in our models.