Deep Neural Network for Cecum Achievement Confirmation During Screening Colonoscopy

Aims According to ESGE guidelines, reaching the cecum is a basic aspect of the quality of screening colonoscopy. An artificial intelligence-based approach can be useful in real-time confirmation of cecal achievement and in monitoring the quality of colonoscopy in an endoscopy unit. We checked the ability of computerized image analysis using deep neural network to confirm reaching the cecum in screening colonoscopy.

Methods Digital database was created, it contained 2671 hand-labeled images from screening colonoscopies collected from more than 250 patients. Among them 2294 is a negative class (without the appendiceal orifice), 377 positive (containing the appendiceal orifice). We randomly divided the database in the ratio of 80 % to 20 % into a training and validation set. Thus, the training base consists of 2136 images, of which 311 are positive and 1825 are negative. The validation set consists of 535 images, including 66 positive images and 469 negative ones.

Results The following results were obtained on a test dataset in the process of the study - the best result on the validation set was AUC = 0.97, and the best value is F1-score = 0.85, when a threshold is th = 0.608. Then the trained model was checked on a test set the area under the curve is equal to AUC = 0.95, F1-score equal 0.9 with a threshold th = 0.462. The average analysis time of one image is 29 ms, which allows to process up to 40 images per second.

Conclusions We have developed and clinically tested an algorithm based on a deep neural network using object classification on endoscopic images to confirm the achievement of the cecum with a high result. These results can be integrated in a quality control system and will lead to a decrease in the number of subjective medical mistakes during screening colonoscopy.

Citation: Kashin S, Zavyalov D, Kuvaev R et al. OP138 DEEP NEURAL NETWORK FOR CECUM ACHIEVEMENT CONFIRMATION DURING SCREENING COLONOSCOPY. Endoscopy 2021; 53: S57.

Publication History

Article published online:
19 March 2021