Development Of Pathology Detection Algorithms Based On Convolutional Neural Networks (CNN)

Aims Development and research of algorithms for detecting pathologies in endoscopic images of stomach based on different architectures of CNN.

Methods As the basis for the detection algorithms two efficient CNN architectures were chosen: Single Shot Detector (SSD) and RetinaNet. To create the database for the algorithms training and testing 54 videos of endoscopic examinations were used. From endoscopic videos every fifth frame was selected. Collected database of endoscopic images included 5942 frames. Due to small size of the endoscopic images database we additionally used pre-training of CNN on images from ImageNet database and data augmentation. For evaluating the quality of the algorithms we used AP (Average Precision) and mAP (Mean Average Precision) as ones of the key metrics for analyzing the quality of the object detectors.

Results All collected images were annotated and divided into three classes: early gastric cancer (902 images), advanced cancer (297 images), benign lesions: intestinal metaplasia, adenoma, hyperplastic polyp, erosion, ulcer, foveolar hyperplasia, xanthoma (1772 images). The database was divided into training (5594 frames) and test (348 frames) datasets with images belonging to different patients. Algorithms for detecting pathologies in endoscopic images based on CNN SSD and RetinaNet were developed, trained and tested on our database. The AP metric values calculated for the developed algorithms for various classes are summarized in [Tab. 1]. The mAP metric value was 0.771 for the CNN SSD and 0.808 for the RetinaNet algorithm.

Tab. 1

CNN used as the basis of the algorithm	AP «cancer»	AP «early cancer»	AP «other pathology»
SSD	0.642	0.937	0.453
RetinaNet	0.873	0.976	0.524

Conclusions Testing has shown that the algorithm based on RetinaNet outperforms the algorithm based on SSD on the AP metric for all classes covered in this study. The mAP metric value for the RetinaNet algorithm was also higher than for the SSD (by 0.037). The obtained values of the metrics are high for both algorithms and prove the possibility of using CNN for detecting pathologies in endoscopic images.

Citation Khryashchev V, Kashin S, Merkulova A et al. eP478 DEVELOPMENT OF PATHOLOGY DETECTION ALGORITHMS BASED ON CONVOLUTIONAL NEURAL NETWORKS (CNN). Endoscopy 2021; 53: S254.

Publication History

Article published online:
19 March 2021