A Hybrid Model for Enhancing 3-D Medical Image Segmentation With Self-Attention Volumetric Transformer and Convolutional Neural Network
IEEE Access · 2026 · 14 · 58619-58634
DOI: 10.1109/access.2026.3680872 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ankit Srivastava und 4 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
In modern healthcare, accurate diagnosis, treatment planning, and surgical guidance are affected mainly by precise segmentation of 3D medical images. Clinicians can more precisely identify complex anatomical structures and detect abnormalities in 3D medical images than in 2D images. Most segmentation work is based on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). CNNs are good at capturing local image details, but struggle to represent broader spatial contexts, whereas ViTs excel at finding long-range dependencies within an image but require substantial processing power. This paper presents a hybrid segmentation approach that combines the strengths of both paradigms. This study employs a Volumetric Transformer Encoder (VN) to capture non-local dependencies across 3D space. The encoder effectively models long-range contextual relationships within volumetric data. A CNN-based decoder is then used to reconstruct fine-grained segmentation maps. This decoder enhances spatial detail and preserves structural fidelity in the final output. The proposed model also uses hierarchical attention and skip connections, which support multiscale learning and maintain semantic consistency. The proposed model is evaluated using the BraTS 2020 benchmark dataset. The results demonstrate competitive performance compared to existing state-of-the-art models. Specifically, the model achieves a higher average Dice score and a lower Hausdorff Distance.
Zitieren
APA
Srivastava, A., Rajak, S., Trivedi, M. C., Saha, A., & Chakraborty, S. (2026). A Hybrid Model for Enhancing 3-D Medical Image Segmentation With Self-Attention Volumetric Transformer and Convolutional Neural Network. IEEE Access, 14, 58619-58634. https://doi.org/10.1109/access.2026.3680872
Vancouver
Srivastava A, Rajak S, Trivedi MC, Saha A, Chakraborty S. A Hybrid Model for Enhancing 3-D Medical Image Segmentation With Self-Attention Volumetric Transformer and Convolutional Neural Network. IEEE Access. 2026;14:58619-58634. doi:10.1109/access.2026.3680872
BibTeX
@article{Srivastava2026,
author = {Srivastava, Ankit and Rajak, Santosh and Trivedi, Munesh Chandra and Saha, Ashim and Chakraborty, Sanjoy},
title = {A Hybrid Model for Enhancing 3-D Medical Image Segmentation With Self-Attention Volumetric Transformer and Convolutional Neural Network},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {58619--58634},
doi = {10.1109/access.2026.3680872}
}
EUVIMED