A Preliminary Study Assessing a Transfer Learning Approach to Intestinal Image Analysis to Help Determine Treatment Response in Canine Protein-Losing Enteropathy

Abstract

Dogs with protein-losing enteropathy (PLE) caused by inflammatory enteritis, intestinal lymphangiectasia, or both, have a guarded prognosis, with death occurring as a result of the disease in approximately 50% of cases. Although dietary therapy alone is significantly associated with a positive outcome, there is limited ability to differentiate between food-responsive (FR) PLE and immunosuppressant-responsive (IR) PLE at diagnosis in dogs. Our objective was to determine if a transfer learning computational approach to image classification on duodenal biopsy specimens collected at diagnosis was able to differentiate FR-PLE from IR-PLE. This was a retrospective study using paraffin-embedded formalin-fixed duodenal biopsy specimens collected during upper gastrointestinal tract endoscopy as part of the diagnostic investigations from 17 client-owned dogs with PLE due to inflammatory enteritis at a referral teaching hospital that were subsequently classified based on treatment response into FR-PLE (n = 7) or IR-PLE (n = 10) after 4 months of follow-up. A machine-based algorithm was used on lower magnification and higher resolution images of endoscopic duodenal biopsy specimens. Using the pre-trained Convolutional Neural Network model with a 70/30 training/test ratio for images, the model was able to differentiate endoscopic duodenal biopsy images from dogs with FR-PLE and IR-PLE with an accuracy of 83.78%. Our study represents an important first step toward the use of machine learning in improving the decision-making process for clinicians with regard to the initial treatment of canine PLE.

Keywords

• machine learning
• canine
• gastrointestinal
• diarrhea