Machine Learning of Feline GI Disorders Using Abdominal Ultrasound Images

Principal Investigator: Parminder Basran

Co-PI: Ian Porter

Department of Clinical Sciences
Sponsor: Cornell Feline Health Center
Title: Machine Learning of Feline GI Disorders Using Abdominal Ultrasound Images
Project Amount: $56,257
Project Period: July 2020 to June 2021

DESCRIPTION (provided by applicant): 

Ultrasound is regularly used as a ‘first-line’ imaging modality for cats with suspicious abdominal disease. Differentiating between inflammatory bowel disease (IBD), small-cell lymphosarcomas (lymphomas), and other abdominal disorders with ultrasound alone is challenging since presentation, history, physical examination, and ultrasound findings can be identical. While ultrasound might be sensitive to an abnormality, it is not very specific. Biopsies in the GI tract are often required to confirm disease; however, biopsies themselves pose challenges, such risk of morbidity from the procedure, deciding between partial thickness endoscopic biopsy vs. full thickness surgical biopsy, and owner convenience, time, and cost. There is a need for robust and sensitive diagnostics in feline abdominal disorders. Artificial Intelligence, or computer algorithms which can detect patterns and trends in large datasets, show promise in addressing many challenges in healthcare. Despite wide adoption in human health, there are only few studies exploring the use of AI in veterinary medicine. To date, there are no published works exploring the use of AI in ultrasound imaging in feline health. The objective of this work is to adopt AI, or more specifically supervised machine learning from image biomarkers (Radiomics), to assist in the diagnosis of abdominal disorders in cats. We will retrospectively obtain 200 transverse ultrasound images of the abdomen (small intestines) and pathology-confirmed diagnoses of IBD, and/or lymphoma. We will manually contour the small intestines and extract image biomarkers from within that region of interest. Then, we will use differences in expressed image biomarkers to train a machine learning classifier. We will compare the machine learning model predictions with the standard approach of classifying disease from subjective appearance and manual measurements, and pathology findings. The sensitivity and sensitivity of the supervised machine learning platform will be calculated and compared with conventional methods of classifying disease. This work will determine if Radiomics could be used to distinguish cats with IBD, lymphoma, and/or other disease without a biopsy.