СЕГМЕНТАЦИЯ МУЛЬТИФАЗНЫХ КТ-ИЗОБРАЖЕНИЙ С ИСПОЛЬЗОВАНИЕМ АНСАМБЛЯ МОДЕЛЕЙ RESUNET

Александр Валерьевич Черемискин; Ирина Леонидовна Каширина

doi:10.17308/sait/1995-5499/2025/3/140-152

Authors

Aleksandr V. Cheremiskin Voronezh State University https://orcid.org/0009-0000-4402-5101 (unauthenticated)
Irina L. Kashirina Russian Technical University MIREA, Voronezh State University https://orcid.org/0000-0002-8664-9817 (unauthenticated)

DOI:

https://doi.org/10.17308/sait/1995-5499/2025/3/140-152

Keywords:

segmentation, ResUNet, Vision Transformer, model ensembles, computed tomography, contrast phases, medical images

Abstract

Automated segmentation of anatomical structures in multiphase computed tomography (CT) images is a critically important task for successful clinical diagnosis. This is especially relevant for the early detection and precise localization of tumors, which allows for timely initiation of treatment and improves patient recovery prospects. The relevance of research in this area is due to the complexities arising from the high degree of variability of structures across different contrast phases. Direct channel concatenation can lead to excessive model complexity and significant computational costs, while standard ensemble methods often ignore non-obvious spatial interrelationships between contrast phases. Thus, the scientific novelty lies in the development of a new Vision Transformer (ViT)-based ensemble approach capable of incorporating contextual information. The proposed method integrates predictions from ResUNet models trained on different contrast phases. The following results were obtained during the study: the best individual ResUNet models achieved a Dice metric of 0.64. Simple averaging of predictions yielded only a small improvement to Dice = 0.66. In contrast, the proposed ViT-based ensemble approach showed significantly higher results — Dice = 0.80. T he obtained results confirm the effectiveness of using transformers for model ensembling, and demonstrate their ability to account for global context. Although the method was tested on a dataset of hepatocellular carcinoma of the liver, the proposed ensemble architecture can be applied in other areas requiring intelligent fusion of model outputs.

Author Biographies

Aleksandr V. Cheremiskin, Voronezh State University

First year postgraduate student of the Department of Mathematical Methods of Operations Research
Irina L. Kashirina, Russian Technical University MIREA, Voronezh State University

Doctor of Engineering Sciences, Professor, Professor of the Department of Mathematical Methods of Operations Research of the Voronezh State University; Professor of the Department of Artificial Intelligence Technologies of the Russia Technical University MIREA

References

Ronneberger O. U-Net: Convolutional Networks for Biomedical Image Segmentation / O. Ronneberger, P. Fischer, T. Brox // Medical Image Computing and Computer-Assisted In-tervention – MICCAI 2015. – Cham : Springer, 2015. – Vol. 9351 (Lecture Notes in Computer Science). – P. 234–241. – DOI: 10.1007/978-3319-24574-4_28

Zhang Z. Road Extraction by Deep Residual U-Net / Z. Zhang, Q. Liu, Y. Wang // IEEE Geo-science and Remote Sensing Letters. – 2018. – Т. 15, № 5. – P. 749–753. – DOI: 10.1109/ LGRS.2018.2802944

Куликов А. А. Сегментация объемных образований печени на мультифазных КТ- изображениях с использованием фреймворка nnU-Net / А. А. Куликов, И. Л. Каширина, Е. Ф. Савкина // Моделирование, оптимизация и информационные технологии. – 2025. – Т. 13, № 1. – DOI: 10.26102/23106018/2025.48.1.040.

Multi-target and multi-stage liver lesion segmentation and detection in multi-phase computed tomography scans [Электронный ресурс] / A. F. Al-Battal [и др.]. – 2024. – Режим доступа: https://arxiv.org/abs/2404.11152. – DOI: 10.48550/arXiv.2404.11152. (Дата обращения: 19.06.2025).

Medical image segmentation based on multi-modal convolutional neural network: Study on image fusion schemes / Z. Guo [et al.] // 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018). – New York: IEEE, 2018. – С. 903–907. – DOI: 10.1109/ISBI.2018.8363717.

Four-phase CT lesion recognition based on multi-phase information fusion framework and spatiotemporal prediction module / S. Qiao [et al.] // BioMedical Engineering OnLine. – 2024. – Т. 23, 103. – DOI: 10.1186/s12938-024-01297-x.

Two-layer Ensemble of Deep Lear ning Models for Medical Image Segmentation / T. Dang [et al.] // Cogn Comput. – 2024. – Т. 16. – P. 1141–1160. – DOI: 10.1007/s12559024-10257-5.

Khoong W. H. BUSU-Net: An Ensemble U-Net Framework for Medical Image Segmentation [Электронный ресурс] / W. H. Khoong. – 2020. – Режим доступа: https://arxiv.org/ abs/2003.01581. – DOI: 10.48550/arXiv.2003.01581. (Дата обращения: 19.06.2025).

UNETR: Transformers for 3D Medical Image Segmentation / A. Hatamizadeh [и др.] // 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). – Waikoloa : IEEE, 2022. – P. 1748–1758. – DOI: 10.1109/ WACV51458.2022.00181.

WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data / K. Bartnik [et al.] // Radiology: Artificial Intelligence. – 2024. – Т. 6, № 6, ст. e240296. – DOI: 10.1148/ ryai.240296.

TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images / J. Wasserthal [и др.] // Radiology: Artificial Intelligence. – 2023. – Т. 5, № 5, ст. e230024. – DOI: 10.1148/ryai.230024.

Salehi S. S. M. Tversky Loss Function for Image Segmentation Using 3D Fully Convolutional Deep Networks / S. S. M. Salehi, D. Erdogmus, A. Gholipour // Machine Learning in Medical Imaging / под ред. Q. Wang [и др.]. – Cham : Springer International Publishing, 2017. – P. 379387. – DOI: 10.1007/978-3-319-67389-9_44.

The Medical Segmentation Decathlon / M. Antonelli [et al.] // Nature Communications. – 2022. – Т. 13, ст. 4128. – DOI: 10.1038/ s41467-022-30695-9.

Современные методы оптимизации и особенности их применения / С. М. Бекетов, Д. А. Зубкова, А. М. Гинцяк [и др.] // Российский технологический журнал. – 2025. – Т. 13, № 4. – С. 78–94. – DOI 10.32362/2500-316X2025-13-4-78-94.