TY - JOUR

T1 - Методы искусственного интеллекта в сердечно-сосудистой хирургии и диагностика патологии аорты и аортального клапана (обзор литературы)

AU - Блеканов, Иван Станиславович

AU - Ким, Глеб Ирламович

AU - Ежов, Федор Валерьевич

AU - Коваленко, Лев Алексеевич

AU - Ларин, Евгений Сергеевич

AU - Разумилов, Егор Сергеевич

AU - Пугин, Кирилл Витальевич

AU - Дадашов, Мурад Сахиб оглы

AU - Пягай, Виктор Александрович

AU - Шматов, Дмитрий Викторович

PY - 2024/7/11

Y1 - 2024/7/11

N2 - The management of patients with aortic and aortic valve pathology is an extremely relevant task. The main problem of this pathology is the absence of obvious symptoms before the onset of a life-threatening condition, dissection or rupture of the aorta. Early timely diagnosis becomes the most relevant in this situation, and imaging research methods play a leading role in this regard. However, the main limiting factor is the speed and quality of image evaluation. Therefore, an actual task is to develop an AI-based physician assistant for image mining (Computer vision, CV). This article provides an overview of modern neural network methods for effective analysis of diagnostic images (MSCT and MRI) relevant for the study of diseases of the cardiovascular system in general and the aorta in particular. One of the main focuses of this analysis is the study of the applicability of modern neural network methods based on the Transformer architecture or the Attention Mechanism, which show high accuracy rates in solving a wide range of tasks in other subject areas, and have a high potential of applicability for qualitative analysis of diagnostic images. An overview of two fundamental problems of image mining is given: classification (ResNet architecture, ViT architect, Swin Transformer architect) and semantic segmentation (2D approaches - U-Net, TransUNet, Swin-Unet, Segmenter and 3D approaches - 3D-Unet, Swin UNETR, VT-UNET). The described methods, with proper fine tuning and the right approach to their training, will effectively automate the process of diagnosing aortic and aortic valve pathology. For the successful implementation of AI development projects, a number of limitations should be taken into account: a high-quality data set, server graphics stations with powerful graphics cards, an interdisciplinary expert group, prepared scenarios for testing in conditions close to real ones.

AB - The management of patients with aortic and aortic valve pathology is an extremely relevant task. The main problem of this pathology is the absence of obvious symptoms before the onset of a life-threatening condition, dissection or rupture of the aorta. Early timely diagnosis becomes the most relevant in this situation, and imaging research methods play a leading role in this regard. However, the main limiting factor is the speed and quality of image evaluation. Therefore, an actual task is to develop an AI-based physician assistant for image mining (Computer vision, CV). This article provides an overview of modern neural network methods for effective analysis of diagnostic images (MSCT and MRI) relevant for the study of diseases of the cardiovascular system in general and the aorta in particular. One of the main focuses of this analysis is the study of the applicability of modern neural network methods based on the Transformer architecture or the Attention Mechanism, which show high accuracy rates in solving a wide range of tasks in other subject areas, and have a high potential of applicability for qualitative analysis of diagnostic images. An overview of two fundamental problems of image mining is given: classification (ResNet architecture, ViT architect, Swin Transformer architect) and semantic segmentation (2D approaches - U-Net, TransUNet, Swin-Unet, Segmenter and 3D approaches - 3D-Unet, Swin UNETR, VT-UNET). The described methods, with proper fine tuning and the right approach to their training, will effectively automate the process of diagnosing aortic and aortic valve pathology. For the successful implementation of AI development projects, a number of limitations should be taken into account: a high-quality data set, server graphics stations with powerful graphics cards, an interdisciplinary expert group, prepared scenarios for testing in conditions close to real ones.

KW - artificial intelligence

KW - cardiovascular surgery

KW - deep learning

KW - diagnosis of aortic pathology

KW - diagnostic images

KW - image classification

KW - image segmentation

KW - neural networks

UR - https://www.sibjcem.ru/jour/article/view/2328

UR - https://www.mendeley.com/catalogue/c1aaaf9f-129d-306f-acee-58fa91de73fc/

U2 - 10.29001/2073-8552-2024-39-2-36-45

DO - 10.29001/2073-8552-2024-39-2-36-45

M3 - статья

VL - 39

SP - 36

EP - 45

JO - Siberian Journal of Clinical and Experimental Medicine

JF - Siberian Journal of Clinical and Experimental Medicine

SN - 2713-2927

IS - 2

ER -