P099 Evaluation of endoscopic mayo score with an artificial intelligence algorithm
Kani, H.T.(1);Ergenc, I.(1);Polat, G.(2);Ozen Alahdab, Y.(1);Temizel, A.(2);Atug, O.(1);
(1)Marmara University- School of Medicine, Gastroenterology, Istanbul, Turkey;(2)Middle East Technical University, Electrical and Electronic Engineering, Ankara, Turkey
Background
Multi-layered convolutional neural networks are artificial intelligence (AI) algorithms that allow to process specific datasets. Endoscopic mayo score (EMS) is an endoscopic scoring tool for ulcerative colitis (UC) that is widely using for evaluating the disease activity to make a further treatment plan. EMS is an endoscopist-depended subjective tool that varies according to the physician’s experience. In this study, our aim was to create a high accuracy EMS diagnostic algorithm to minimize endoscopist-depended inconsistency and standardize the patient care.
Methods
We collected the endoscopic images of UC patients between December 2011 and July 2019 from electronic database of our gastroenterology institute. Images with insufficient bowel cleaning, artifact, retroflection images, terminal ileum images and pouch patients were excluded. Two blinded gastroenterologists evaluated and tagged the images according to the EMS. Images with a disagreement were excluded for a further evaluation. AI algorithm was performed with Python by using PyTorch library. The dataset was divided into two (85% was used for training and %15 was used for test). ResNet18 model was also used for training.
Results
A total of 19690 images of 572 patients from 1053 colonoscopies were identified for the study. The mean procedure number was 1.8 per patient and the mean image number was 18.7 for per colonoscopy. Four thousand and six hundred images without any disagreement between two gastroenterologists were included to the analysis. Two thousand eight hundred and thirteen (61.65%) images were tagged as EMS 0, 956 (20.66%) were tagged as EMS 1, 406 (8.77%) were tagged as EMS 2 and 413 (8.92%) were tagged as EMS 3. Accuracy was found 73.16% with a sensitivity of 773.2% and specifity of 92.9% in assessment of all EMS groups (Image 1). Also, the accuracy of severe mucosal disease diagnosis (EMS 0 and 1 vs EMS 2 and 3) was 96.3% with a sensitivity of 98.2% and specifity of 86.5% (Image 2) with a perfect reproductivity (к: 1.00). The performance of the remission diagnosis (EMS 0 vs EMS 1,2 and 3) was done with a 92% accuracy.
Conclusion
This is an ongoing study and the preliminary results of our EMS diagnosis algorithm was promising with a high accuracy. The accuracy and sensitivity would be improved by including more images and improving the algorithm. The use of AI in daily IBD practice can eliminate the subjectivity according to the endoscopist in diagnosis and assessing the disease severity for treatment decision.