TAU-Net: Mitigating Oversegmentation with a Topology and CBAM Augmented Attention U-Net.

Saved in:

Bibliographic Details
Title:	TAU-Net: Mitigating Oversegmentation with a Topology and CBAM Augmented Attention U-Net.
Authors:	Kumar, Sanjeet¹ sanjuonline1@gmail.com, Pilania, Urmila² urmilapilania@gmail.com
Source:	IAENG International Journal of Computer Science. Jul2026, Vol. 53 Issue 7, p2727-2739. 13p.
Subjects:	Topology, Image segmentation, Diagnostic imaging, Deep learning, Loss functions (Statistics), Magnetic resonance imaging
Abstract:	Deep Learning (DL) models are widely used in medical imaging, particularly with MRI data, due to their strong capability in capturing complex spatial features and achieving high volumetric accuracy. However, these models often generate topological errors such as disconnected tumor regions, holes, and anatomically inconsistent boundaries, primarily because they optimize voxel-level losses without enforcing structural constraints. To address this limitation, this paper proposes TAU-Net, an advanced architecture that integrates topological constraints into the learning process. TAU-Net utilizes Contrast-Enhanced T1-weighted (T1ce) and FLAIR MRI modalities to capture complementary information. It extends Attention U-Net with two key enhancements: Convolutional Block Attention Modules (CBAM) for improved feature refinement, and a hybrid loss function combining Dice, Focal, and a persistent homology-based topological loss to preserve global structural properties. Evaluated on the BraTS2021 dataset, TAU-Net achieves strong segmentation performance with Dice scores of 0.912 (whole tumor), 0.831 (tumor core), and 0.773 (enhancing tumor), while improving topological consistency by reducing erroneous connected components by ~40%. It also attains a Betti Number Error of 1.7 and a Defect Count of 2.5. The results demonstrate that incorporating topological loss enhances structural accuracy, reduces implausible predictions, and improves the anatomical and clinical reliability of tumor segmentation. [ABSTRACT FROM AUTHOR]
	Copyright of IAENG International Journal of Computer Science is the property of International Association of Engineers (IAENG) and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database:	Engineering Source

Description
Abstract:	Deep Learning (DL) models are widely used in medical imaging, particularly with MRI data, due to their strong capability in capturing complex spatial features and achieving high volumetric accuracy. However, these models often generate topological errors such as disconnected tumor regions, holes, and anatomically inconsistent boundaries, primarily because they optimize voxel-level losses without enforcing structural constraints. To address this limitation, this paper proposes TAU-Net, an advanced architecture that integrates topological constraints into the learning process. TAU-Net utilizes Contrast-Enhanced T1-weighted (T1ce) and FLAIR MRI modalities to capture complementary information. It extends Attention U-Net with two key enhancements: Convolutional Block Attention Modules (CBAM) for improved feature refinement, and a hybrid loss function combining Dice, Focal, and a persistent homology-based topological loss to preserve global structural properties. Evaluated on the BraTS2021 dataset, TAU-Net achieves strong segmentation performance with Dice scores of 0.912 (whole tumor), 0.831 (tumor core), and 0.773 (enhancing tumor), while improving topological consistency by reducing erroneous connected components by ~40%. It also attains a Betti Number Error of 1.7 and a Defect Count of 2.5. The results demonstrate that incorporating topological loss enhances structural accuracy, reduces implausible predictions, and improves the anatomical and clinical reliability of tumor segmentation. [ABSTRACT FROM AUTHOR]
ISSN:	1819656X