ACS-SegNet: An Attention-Based CNN-SegFormer Segmentation Network for Tissue Segmentation in Histopathology
This repository contains the implementation details of the ACS-SEGNET model.
Our paper preprint is available on arXiv: https://arxiv.org/abs/2510.20754
BibTex entry:
@article{torbati2025acs,
title={ACS-SegNet: An Attention-Based CNN-SegFormer Segmentation Network for Tissue Segmentation in Histopathology},
author={Torbati, Nima and Meshcheryakova, Anastasia and Woitek, Ramona and Mechtcheriakova, Diana and Mahbod, Amirreza},
journal={arXiv preprint arXiv:2510.20754},
year={2025}
}
Table 1. Segmentation results on the GCPS dataset
| Method | μIoU (%) | μDice (%) |
|---|---|---|
| DGAUNet [15] | 75.95 ± 0.20 | 86.33 ± 0.13 |
| SegFormer [13] | 70.90 ± 0.38 | 82.97 ± 0.26 |
| ResNetUNet [17] | 75.65 ± 0.08 | 86.13 ± 0.05 |
| TransUNet [6] | 74.84 ± 0.10 | 85.61 ± 0.09 |
| CS-SegNet | 76.68 ± 0.15 | 86.80 ± 0.06 |
| ACS-SegNet | 76.79 ± 0.14 | 86.87 ± 0.09 |
Table 2. Segmentation results on the PUMA dataset
| Method | μIoU (%) | μDice (%) |
|---|---|---|
| DGAUNet [15] | 44.35 ± 1.76 | 53.69 ± 1.91 |
| SegFormer [13] | 46.78 ± 2.58 | 58.25 ± 2.01 |
| ResNetUNet [17] | 58.42 ± 1.98 | 71.58 ± 1.21 |
| TransUNet [6] | 62.43 ± 2.47 | 74.63 ± 1.91 |
| CS-SegNet | 63.67 ± 1.23 | 75.55 ± 1.71 |
| ACS-SegNet | 64.93 ± 2.28 | 76.60 ± 1.36 |
This project has been conducted through a joint WWTF-funded project (Grant ID: 10.47379/LS23006) between the Medical University of Vienna and Danube Private University.
Academic Research Use: This work is provided "as is", without warranty of any kind.