| Title: |
MAAR-Net: Multi-scale attention-assisted residual neural network for renal microvascular structure segmentation. |
| Authors: |
Tingting Wang; Baoguang Lin; Tong Jiang; Hengjiao Wang; Defu Yang; Feng Shang; Long Li; Ying Li; Mengyan Zhao; Ying Xu; Ying Yan |
| Source: |
PLoS ONE, Vol 21, Iss 3, p e0342752 (2026) |
| Publisher Information: |
Public Library of Science (PLoS) |
| Publication Year: |
2026 |
| Collection: |
Directory of Open Access Journals: DOAJ Articles |
| Subject Terms: |
Medicine; Science |
| Description: |
Renal disease represents a significant public health concern, with renal microvascular lesions playing a crucial role in disease progression. Accurate segmentation of this microvasculature is therefore essential for precise pathologic evaluation. While deep learning offers substantial opportunities in medical image segmentation, the complex structure of renal microvessels poses a considerable challenge. Existing models often struggle to achieve high segmentation accuracy while maintaining branch continuity, suppressing background interference, and delineating tissue boundaries. To address these challenges, we propose a novel deep learning architecture termed the Multiscale Attention-Assisted Residual Neural Network (MAAR-Net). Built upon a U-Net encoder-decoder backbone, MAAR-Net integrates multiscale residual blocks and a high-semantic feature extraction layer to expand the receptive field and enrich semantic information. Depth-separable convolutional attention blocks are incorporated into skip connections to enhance the capture of global and local features, thereby refining segmentation performance. Additional segmentation branches are included to aggregate multi-receptive-field information, further improving segmentation efficiency. Our experiments, conducted on the HuBMAP dataset of 2D PAS-stained kidney histology images, demonstrate the effectiveness of MAAR-Net. The model achieves an Intersection over Union (IoU) of 0.5063 and an F1-score of 0.6751, outperforming other mainstream segmentation models. To facilitate clinical deployment, the optimized model is subsequently compressed via structured pruning to reduce size and increase speed, followed by quantification to lower computational resource consumption. These optimizations ensure the model's suitability for real-time performance in practical diagnostic applications, independent of dedicated workstations or cloud servers. The results collectively validate the robustness and practical utility of our approach for accurate renal microvessel segmentation ... |
| Document Type: |
article in journal/newspaper |
| Language: |
English |
| Relation: |
https://doi.org/10.1371/journal.pone.0342752; https://doaj.org/toc/1932-6203; https://doaj.org/article/bfadbcea09a04093bfdb6b6cc6ea70fa |
| DOI: |
10.1371/journal.pone.0342752 |
| Availability: |
https://doi.org/10.1371/journal.pone.0342752; https://doaj.org/article/bfadbcea09a04093bfdb6b6cc6ea70fa |
| Accession Number: |
edsbas.26E5F2D7 |
| Database: |
BASE |