Year: 2025
Type: article
Abstract: АНАЛІЗ МЕТОДІВ СЕГМЕНТАЦІЇ БІОМЕДИЧНИХ 3D-ДАНИХ ЗА ДОПОМОГОЮ ГЛИБОКОГО НАВЧАННЯLashchevska N.O., Cherevyk O.V. Analysis of biomedical 3D data segmentation methods using deep learning.This article provides a comprehensive review of modern approaches to the segmentation of biomedical 3D scans using deep learning techniques.A comparative analysis of neural network architectures is conducted, particularly convolutional neural networks (CNNs), Transformers, and the latest Mamba-type models, which offer varying levels of computational complexity, accuracy, and generalization capability.The advantages of hybrid architectures, multimodal approaches, and pretraining strategiesincluding self-supervised learning and generative adversarial networks (GANs) -are described as methods to improve performance under conditions of limited labeled data.The article highlights current segmentation challenges related to the high complexity of biomedical images, variability in scanning protocols, and resource constraints.Practical recommendations are offered for selecting the appropriate architecture based on task requirements, dataset characteristics, and available computational resources.Standard evaluation metrics (e.g., Dice coefficient, Hausdorff distance) and the use of publicly available medical datasets (BraTS, LiTS, ACDC) for model benchmarking are also discussed.
Authors: Н. О. Лащевська, О. В. Черевик
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W4412845791