A deep learning framework for glomeruli segmentation with boundary attention
Authors: Behnaz Elhaminia, Catherine King, Jiaqi Lv, Lorraine Harper, Paul Moss, Owen Cain, Dimitrios Chanouzas, Shan E Ahmed Raza
Organizations: Tissue Image Analytics (TIA) Centre, Dept. of Computer Science, University of Warwick, UK · Dept. of Immunology and Immunotherapy, University of Birmingham, UK · Renal Unit, Queen Elizabeth Hospital Birmingham, UHB NHS Foundation Trust, UK · School of Applied Health Sciences, University of Birmingham, UK · Dept. of Cellular Pathology, Queen Elizabeth Hospital Birmingham, UK
Accurate detection and segmentation of glomeruli in kidney tissue are essential for diagnostic applications. Traditional deep learning methods primarily rely on semantic segmentation, which often fails to precisely delineate adjacent glomeruli. To address this challenge, we propose a novel glomerulus detection and segmentation model that emphasises boundary separation. Leveraging pathology foundation models, the proposed U-Net-based architecture incorporates a specialised attention decoder designed to highlight critical regions and improve instancelevel segmentation. Experimental evaluations demonstrate that our approach surpasses state-of-the-art methods in both Dice score and Intersection over Union, indicating superior performance in glomerular delineation.
Figures & tables
Figure 1: An illustration of the architecture and key processing stages. Deep features extracted by the Virchow2 foundation model are concatenated with the input image and fed into the encoder. The attention module exploits these features to extract critical boundary-related information, which helps the model segment boundaries with greater precision and focus.
Figure 2: Comparison of static and adaptive weight maps. (A) Glomeruli mask, (B) Static distance-transform weight map, and (C) Adaptive boundary weighting map. The adaptive weights assign higher importance to challenging boundaries between closely adjacent glomeruli.
HubMap
REACTIVAS
Method
PAS
HE
PAS
All
Proposed
0.9630
0.89±0.3
0.86±0.1
0.88 ± 0.01
Eff - UNet
0.9111
0.84±0.2
0.83±0.4
0.84±0.3
Tom
0.9515
0.80±0.1
0.82±0.4
0.81±0.2
Gleb
0.9507
-
-
-
WGO
0.9503
0.79±0.2
0.81±0.3
0.80±0.2
Table 1: Performance Comparison across two datasets in terms of Dice scores ( μ±σ ). For REACTIVAS, we add the results for HE and PAS separately for further insights. Training code is not available for Gleb. Therefore, we could not train Gleb on the REACTIVAS dataset.
HubMap
REACTIVAS
Method
PAS
HE
PAS
All
Proposed
0.941
0.88 ± 0.2
0.86 ± 0.3
0.87 ± 0.4
Eff - UNet
0.904
0.82± 0.1
0.81 ± 0.1
0.82 ± 0.3
Tom
0.943
0.78 ± 0.2
0.79 ± 0.4
0.78 ± 0.3
WGO
0.931
0.77 ± 0.2
0.79 ± 0.1
0.78 ± 0.6
Table 2: Performance Comparison across two datasets in terms of IoU ( μ±σ ). For REACTIVAS, we add the results for HE and PAS separately for further insights. IoU results were not available for the Gleb model.
Figure 3: Comparison of glomerulus segmentation results, with PAS images in the top row and HE images in the bottom row. (A) Ground truth, (B) Proposed model, (C) Eff-UNet, (D) TOM, and (E) WGO. The patches are taken from the test set of the REACTIVAS dataset and illustrate challenging cases with closely spaced glomeruli.
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Interdisciplinary Institute for Medical Engineering, Fuzhou University, Fuzhou, China · Medical Optical Technology R&D Center, Research Institute of Tsinghua, Pearl River Delta, Guangzhou, China +6
Department of Mechanical Engineering, Yonsei University, Seoul, 03722, South Korea · Institute of Vision Research, Department of Ophthalmology, Severance Hospital, Yonsei University, Seoul, 03722, South Korea