THE USE OF THE EFFICIENTDET NEURAL NETWORK IN THE TASK OF DETECTING STOMACH PATHOLOGIES ON VIDEO IMAGES OF ENDOSCOPIC EXAMINATION

Abstract

Background. Obtaining the most reliable information when performing endoscopic examinations of the gastrointestinal tract makes it possible to detect and classify dangerous pathologies, including oncological ones, at an early stage. This significantly reduces mortality among patients with these types of diseases. The problem of detecting stomach pathologies on the video data of endoscopic studies is considered. Materials and methods. The analysis of the effectiveness of using a fairly new architecture of the EfficientDet convolutional neural network for detecting dangerous stomach pathologies on video images of endoscopic examinations is carried out. 54 videos of stomach studies conducted in the endoscopic department of the Yaroslavl Regional Clinical Oncology Hospital were used to train and test deep machine learning algorithms. Results. The results of the developed algorithm are analyzed in comparison with the popular approach based on the architecture of the convolutional neural network SSD300. The dependences of the values of the loss functions of the classification subnet and the regression subnet calculated on the training sample, as well as the values of the standard metrics F1, mAP, Precision and Recall are obtained. The achievement of a significant superiority in the metric of average accuracy is shown while maintaining good indicators according to the criteria of robustness and processing speed of the video frame. Conclusions. The developed algorithm can be used as the main one in the implementation of a neural network module for detecting pathologies on endoscopic images of the stomach.

Keywords

• endoscopic video data
• machine learning
• convolutional neural network
• object detection
• stomach pathology
• quality metrics