Development of the Expert AI System. Neural Networks and Pathology of the Thoracic Aorta

P0101 

Abstract Submission 

Gleb Kim (1), Ivan Blekanov (2), Murad Dadashov (3)

(1) N/A, Russia, (2) Saint Petersburg State University, Saint Petersburg , NA, (3) Saint Petersburg State University, Saint Petersburg, NA

Gleb Kim    -  Contact Me
N/A

Ivan Blekanov    -  Contact Me
Saint Petersburg State University

Murad Dadashov    -  Contact Me
Saint Petersburg State University

Gleb Kim    -  Contact Me
N/A

Objective
An aortic aneurysm is a life-threatening condition that can cause aortic dissection or rupture and most often requires surgical treatment. In order to successfully perform operations on the thoracic aorta, it is necessary to have a specialized Aortic team that allows to perform complex reconstructive operations with minimal complications. Another option is direct contact to experts or telemedicine. In emergency situations, it is not always possible to use such assistance, experts cannot be available around the clock. Thus, one of the solutions to this problem is the use of an expert system based on artificial intelligence technologies.
The aim of study: Analysis of the development of the EXPERT AI System for the examination of thoracic aortic pathology.
Methods
Currently, work is underway to develop the EXPERT AI System. A team of cardiac surgeons and cardiologists from the Saint Petersburg State University Hospital and Data Science specialists from St. Petersburg State University are involved in the development. The system is based on the use of an ensemble of neural networks and the analysis of a large amount of data, including anthropometry, clinical indicators, computed tomography and transthoracic echocardiography. The main technology used in the work is modern models of convolutional neural networks and transfer learning, which are used in the task of segmentation, including medical images. In particular, the work conducted an experiment to assess the quality of three neural network models: a model based on the U-Net architecture with a ResNet-50 encoder, TransUnet and SWIN transformers. The models under study were implemented in the Python 3 programming language, and PyTorch was chosen as the framework. To analyze the images of the aorta and the learning process of neural networks, both data from existing labeled datasets and computed tomography data of the chest and aorta organs of the patients selected and labeled by us were used.
Results
At the moment, 3 neural network models ("U-Net+ResNet-50", TransUnet and SWIN) have been developed and trained for automatic detection of the aorta of the heart on CT scans and methods for constructing its digital 3D model in full size. The resulting digital model of the aorta is planned to be used as a preparatory data processing procedure for neural network methods for segmenting the diameter of the aorta, searching and detailing pathological abnormalities/disorders in the aorta.
Conclusions
The widespread use of artificial intelligence in cardiac surgery is just beginning. However, our team is one of the leaders in this area. The lack of a sufficient number of experts in the field of aortic surgery, as well as the need for assistance in decision-making, is a key problem that can be solved through the use of an expert system.

Ascending Aorta