Endoscopic Image Enhanced Deep Learning Algorithm for Inflammatory Bowel Disease (IBD) Polyp Detection: Feasibility Study.

Gastrointestinal endoscopy is a commonly used diagnostic procedure for surveillance colonoscopies in patients with Inflammatory Bowel Diseases (IBD). Patients with IBD can have benign, inflammatory and or malignant polyps that require further testing and evaluation. Endoscopic image acquisition often comes with the challenges of poor quality, and often image preprocessing steps are essential for developing artificial intelligence (AI) assisted models for improving IBD polyp detection. Through artificial intelligence, detection and differentiation of these polyps can be made more efficient as it eliminates human error. The purpose of this work was to evaluate the utility of several digital filters such as average filter (AF), median filter (MF), gaussian filter (GF) and Savitzky Golay Filter (SG) to enhance the images to improve deep learning model detection of IBD polyps and compare the performance without enhancement. IBD polyp images from high-definition white light endoscopy (HDWLE) from Mayo Clinic, GIH Division were used to develop a You-Only-Look-Once (YOLO) model which employs conventional neural networks (CNN) to detect IBD polyps. Varying filter kernels such as, 3x3, 5x5, 7x7, 9x9, 11x11 and 13x3 were employed for all four filter types and YOLO model was deployed for each case. Performance was measured using precision and recall curve and measuring the area under the curve (AUC). 80% data was used for training and validation and 20% was used for testing. A moderate 5-10% improvement in deep learning model performance was observed. Further testing with different model parameters and filter settings is required to validate these findings.