Multicentre Bi-atrial Segmentation from LGE-MRI for Atrial Fibrillation with a 2D and 3D Framework

Malitha Gunawardhana; Gregory B. Sands; Mark L. Trew; Jichao Zhao

eess.IVSep 18, 2026

Multicentre Bi-atrial Segmentation from LGE-MRI for Atrial Fibrillation with a 2D and 3D Framework

Authors: Malitha Gunawardhana, Gregory B. Sands, Mark L. Trew, Jichao Zhao

Abstract

Accurate delineation of bi-atrial structures from late gadolinium enhancement MRI (LGE-MRI) is an important prerequisite for structural analysis and future fibrosis-quantification workflows in atrial fibrillation (AF). However, automated segmentation is challenging due to thin-walled anatomy, domain shifts across imaging centres, and limited benchmarking of existing methods. This study presents a two-stage segmentation framework and benchmarking platform for evaluating how ROI localisation, encoder design, 2D/3D dimensionality, and ensemble fusion affect bi-atrial wall and cavity segmentation across multicentre LGE-MRI datasets. The framework integrates 3D localisation and fine segmentation using 2D and 3D U-Net variants with ResNeXt encoders and compares them with convolutional, transformer-based, and state-space architectures. { Evaluation across three independent cohorts assessed accuracy and cross-domain transfer without target-domain fine-tuning. Cavity segmentation transferred more consistently across centres than atrial wall segmentation, while wall performance remained sensitive to domain shift, particularly in the Kobe cohort.} By quantifying how 2D, 3D, and ensemble architectures behave across centres and between walls and cavities, this work provides a reproducible benchmark for future methodological development and clinical validation.

Figures & tables

Study / benchmark	Evaluation type	Label focus
LA2018 [ 51 ]	Challenge	LA cavity
LAScarQS2022 [ 55 ]	Challenge	LA cavity, LAScar
MBAS2024 [ 33 ]	Challenge	LA cavity, RA cavity, Bi atrial wall
RAS [ 1 ]	Individual benchmark	RA cavity
biAtriaNet [ 8 ]	Individual evaluation	LA cavity, RA cavity, Bi atrial wall
TASSNet (This study)	Controlled benchmark	LA wall, RA wall, LA cavity, RA cavity

Table 1: Comparison with existing LGE-MRI atrial segmentation benchmarks and related studies. LA: left atrium; RA: right atrium; LAScar: left atrial scar; MBAS: Multi-class Bi-Atrial Segmentation.

Figure 1: Overview of the proposed two-stage segmentation framework for atrial structures from 3D LGE-MRI images. In Stage 1, a 3D U-Net is employed to extract regions of interest (ROI) from the input LGE-MRI volume, concentrating on the atrial structures. The output of this stage provides localised ROI, reducing spatial complexity for more precise segmentation in the following stage. In Stage 2, both 2D and 3D U-Net architectures are used to segment the atrial structures within the extracted ROIs finely. The 2D U-Net processes data slice-by-slice, generating 2D probability maps, while the 3D U-Net uses volumetric information to produce 3D probability maps. The final predicted mask, accurately delineating the left and right atrial walls and cavities, is obtained by ensembling the outputs of both networks.

Dataset / Source	Training Images	Testing Images	Z-axis Slices	Final Resolution (mm)
Utah	80	20	44	0.625 × 0.625 × 2.5
Waikato	–	11	44	0.625 × 0.625 × 2.5
Kobe	–	10	90	0.630 × 0.630 × 1.20

Table 2: Summary of the LGE-MRI datasets used in this study, including dataset size, resolution, and source institution.

Panel of Figure 2 — Figure 2: Utah dataset clustering analysis: (a) silhouette score, (b) K-means elbow curve, (c) Ward hierarchical clustering, and low-dimensional visualisations using (d) PCA, (e) t-SNE, and (f) UMAP.

Dataset	Intensity	Skewness	Kurtosis	Label Ent.	Image Ent.	MI
Utah (Train)	12.26 $\pm$ 19.73	3.33	23.91	0.1046	3.1799	0.0324
Utah (Test)	8.96 $\pm$ 15.92	4.26	36.60	0.0955	2.8530	0.0324
Utah (Combined)	11.92 $\pm$ 19.40	3.43	25.19	0.1037	3.1472	0.0325
Waikato	22.30 $\pm$ 32.69	2.06	5.32	0.0872	3.6704	0.0216
Kobe	28.44 $\pm$ 41.63	1.77	3.62	0.1319	3.0844	0.0348

Table 3: Descriptive statistics for Utah, Waikato, and Kobe datasets. Metrics include intensity (mean

\pm

standard deviation), skewness, kurtosis, label entropy (Ent.), image entropy, and mutual information (MI).

Model	Config	RA Wall			LA wall			RA cavity			LA cavity
	Config	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95
nnUNet	2D	0.738	0.548	2.094	0.594	0.771	3.125	0.915	0.769	3.031	0.918	0.773	3.098
	3D	0.735	0.557	2.069	0.609	0.739	2.778	0.913	0.771	2.786	0.920	0.758	2.866
	3D low	0.723	0.622	2.408	0.574	0.973	4.231	0.908	0.841	3.094	0.917	0.821	3.216
	3D cas.	0.735	0.556	1.981	0.609	0.739	2.797	0.914	0.779	2.751	0.920	0.758	2.916
	Ens.	0.740	0.541	2.040	0.610	0.751	2.856	0.915	0.762	2.817	0.921	0.750	2.949

Table 4: Performance of different methods on the Utah dataset for Right Atria (RA) wall, Left Atria (LA) wall, RA cavity, and LA cavity. DSC: Dice Similarity Coefficient, ASD: Average Surface Distance, HD95: 95th percentile of Hausdorff Distance. The best values are highlighted . 3D low:3D low resolution, 3D cas:3D cascade model, Ens: Ensemble.

Model	Config	RA Wall			LA wall			RA cavity			LA cavity
	Config	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95
nnUNet	2D	0.608	1.724	6.738	0.470	2.098	8.897	0.785	1.847	6.491	0.816	2.542	11.737
	3D	0.667	1.009	4.543	0.560	1.198	5.283	0.846	1.292	4.522	0.867	2.046	10.008
	3D low	0.637	1.100	4.699	0.507	1.891	9.127	0.857	1.282	4.206	0.847	2.662	13.585
	3D cas.	0.669	0.961	4.095	0.547	1.356	6.293	0.862	1.184	4.030	0.864	2.192	11.017
	Ens.	0.676	0.837	3.217	0.559	1.209	5.397	0.863	1.177	3.934	0.866	2.075	10.295

Table 5: Performance of different methods on the Waikato dataset for Right Atria (RA) wall, Left Atria (LA) wall, RA cavity, and LA cavity. DSC: Dice Similarity Coefficient, ASD: Average Surface Distance, HD95: 95th percentile of Hausdorff Distance. The best values are highlighted . 3D low:3D low resolution, 3D cas:3D cascade model, Ens: Ensemble.

Model	Config	RA Wall			LA wall			RA cavity			LA cavity
	Config	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95	DSC	ASD	HD95
nnUNet	2D	0.444	4.710	17.779	0.431	3.681	12.139	0.778	4.079	19.364	0.852	2.384	8.295
	3D	0.502	2.822	13.625	0.440	4.077	12.964	0.800	2.580	9.067	0.837	2.596	9.258
	3D low	0.454	5.735	21.432	0.472	2.971	13.037	0.703	4.660	16.492	0.861	2.570	10.017
	3D cas.	0.481	3.328	14.127	0.457	3.599	12.093	0.790	3.066	10.379	0.795	5.219	27.413
	Ens.	0.464	5.285	18.483	0.436	4.414	15.975	0.753	4.161	12.495	0.789	4.671	20.766

Table 6: Performance of different methods on the Kobe dataset for Right Atria (RA) wall, Left Atria (LA) wall, RA cavity, and LA cavity. DSC: Dice Similarity Coefficient, ASD: Average Surface Distance, HD95: 95th percentile of Hausdorff Distance. The best values are highlighted. 3D low:3D low resolution, 3D cas:3D cascade model, Ens: Ensemble.

Figure 3: Overall performance of TASSNet with ResNeXt encoders across all three datasets, evaluated using 2D, 3D, and ensemble configurations for the left atrial (LA) and right atrial (RA) cavity and wall.

Dataset	Inference Time (s)
	2D	3D
Utah	2.38 $\pm$ 0.68	4.91 $\pm$ 1.33
Waikato	2.51 $\pm$ 0.67	4.71 $\pm$ 1.56
Kobe	5.47 $\pm$ 0.91	9.81 $\pm$ 2.74

Table 7: Average inference time (in seconds) for 2D and 3D models across different datasets, with standard deviation.

Figure 4: Slice-by-slice evaluation of the predicted performance using dice score. Values are averaged across the 2D, 3D and Ensemble models from the TASSNet with ResNeXt. LA- Left Atria, RA- Right Atria.

Figure 5: Comparison of Left Atria (LA) and Right Atria (RA) Wall and Cavity Segmentation from MRI Segmentation results are shown for three datasets: A) Utah, B) Waikato, and C) Kobe. Each column represents a different MRI scan. LA cavity (yellow), RA cavity (blue), LA wall (green), and RA wall (red). Key anatomical structures are labelled: AO – aorta; MV/TCV – mitral/tricuspid valve; PV – pulmonary vein.

Figure 6: Three-dimensional visualisation of the predicted left and right atria (LA and RA) for one scan from each dataset. The first row shows the model predictions, while the second row presents the corresponding error maps overlaid on the ground truth segmentations. Average Dice scores are shown in the bottom left corner of the error maps.

Figure 7: Comparison of segmentation predictions from 2D (first column), 3D (second column), and ensemble model (third column). Each row corresponds to a different LGE-MRI scan. The segmented anatomical structures are colour-coded as follows: left atrial (LA) cavity in yellow, right atrial (RA) cavity in blue, LA wall in green, and RA wall in red. The yellow bounding boxes highlight regions where noticeable differences are observed between 2D and 3D model predictions, illustrating the variability in anatomical boundary delineation across configurations.

Figure 8: Visualisation of failure cases. Each column corresponds to a different LGE-MRI scan. The segmented LA cavity is shown in yellow, RA cavity in blue, LA wall in green, and RA wall in red. Yellow bounding boxes highlight regions of segmentation failure, particularly in atrial wall regions, in comparison to the ground truth.

Metric	Utah		Waikato		Kobe
	LA	RA	LA	RA	LA	RA
MAE	7.098	5.809	9.075	8.948	9.161	15.852
RMSE	8.232	7.496	13.605	10.923	14.634	19.384
MAPE (%)	8.340	7.885	8.578	12.727	9.869	15.713
$\rho$	0.947	0.981	0.984	0.849	0.913	0.818
RC	0.768	1.056	0.717	0.756	0.960	0.666

Table 8: Quantitative assessment of atrial volume measurements (in mL) for the Utah, Waikato, and Kobe datasets, covering both left atrium (LA) and right atrium (RA). Metrics include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Pearson correlation coefficient (

\rho

), Regression Coefficient (RC), and bias.

Metric	Utah		Waikato		Kobe
	LA	RA	LA	RA	LA	RA
MAE	3.855	2.675	6.289	3.853	8.073	5.623
RMSE	4.819	3.344	7.861	4.816	10.091	7.029
MAPE (%)	4.391	3.260	6.114	5.206	8.163	6.356
$\rho$	0.868	0.968	0.849	0.952	0.896	0.753
RC	0.619	0.884	0.456	0.826	1.500	0.484

Table 9: Quantitative evaluation of atrial diameter measurements (in mm) on the Utah, Waikato, and Kobe datasets for both the left atrium (LA) and right atrium (RA). Metrics include Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Pearson correlation coefficient (

\rho

), Regression Coefficient (RC), and bias.

Explore similar work

C2W-Tune: Cavity-to -Wall Transfer Learning for Thin Atrial Wall Segmentation in 3D LGE-MRI
Mar 26, 2026Yusri Al-Sanaani, Rebecca Thornhill, Sreeraman RajanMedical Image Segmentation 3D Medical Image Segmentation
Systems and Computer Engineering, Carleton University, Ottawa, Canada · Department of Radiology, University of Ottawa, Ottawa, Canada
Multi-Stage Bi-Atrial Segmentation Framework from 3D Late Gadolinium-Enhanced MRI using V-Net Family Models
Apr 29, 2026Hao Wen, Jingsu KangImage Segmentation Magnetic Resonance Imaging
College of Science, China Agricultural University, Beijing, China · Tianjin Medical University, Tianjin, China
Few-Shot Left Atrial Wall Segmentation in 3D LGE MRI via Meta-Learning
Mar 26, 2026Yusri Al-Sanaani, Rebecca Thornhill, Pablo Nery +5Medical Image Segmentation 3D Medical Image Segmentation
Department of Systems and Computer Engineering, Carleton University, Ottawa, Canada · Department of Radiology, Radiation Oncology, and Medical Physics, University of Ottawa, Ottawa, Canada · Division of Cardiology, Department of Medicine, University of Ottawa Heart Institute, Ottawa, Canada

Mar 26, 2026J/K move · Enter open · S save

C2W-Tune: Cavity-to -Wall Transfer Learning for Thin Atrial Wall Segmentation in 3D LGE-MRI

Yusri Al-Sanaani, Rebecca Thornhill, Sreeraman Rajan

Systems and Computer Engineering, Carleton University, Ottawa, Canada · Department of Radiology, University of Ottawa, Ottawa, Canada

Medical Image Segmentation 3D Medical Image Segmentation Cardiac MRI Transfer Learning

Open paper AskPDF arXiv

Accurate segmentation of the left atrial (LA) wall in 3D late gadolinium-enhanced MRI (LGE-MRI) is essential for wall thickness mapping and fibrosis quantification, yet it remains challenging due to the wall's thin geometry, complex anatomy, and low contrast. We propose C2W-Tune, a two-stage cavity-to-wall transfer framework that leverages a high-accuracy LA cavity model as an anatomical prior to improve thin-wall delineation. Using a 3D U-Net with a ResNeXt encoder and instance normalization, Stage 1 pre-trains the network to segment the LA cavity, learning robust atrial representations. Stage 2 transfers these weights and adapts the network to LA wall segmentation using a progressive layer-unfreezing schedule to preserve cavity features while enabling wall-specific refinement. On the 2018 LA Segmentation Challenge dataset, C2W-Tune outperformed an architecture-matched baseline trained from scratch. The wall Dice score increased from 0.623 to 0.814, surface Dice at 1 mm increased from 0.553 to 0.731, 95th-percentile Hausdorff distance (HD95) decreased from 2.95 mm to 2.55 mm, and average symmetric surface distance (ASSD) decreased from 0.71 mm to 0.63 mm. Under reduced supervision using 70 training volumes sampled from the same training set, C2W-Tune achieved a Dice of 0.78, remaining competitive with recent multi-class bi-atrial benchmarks, typically 0.6-0.7. These results show that anatomically grounded task transfer with controlled fine-tuning improves accuracy for thin LA wall segmentation in 3D LGE-MRI.

Abstract

Explore similar work

C2W-Tune: Cavity-to -Wall Transfer Learning for Thin Atrial Wall Segmentation in 3D LGE-MRI

Multi-Stage Bi-Atrial Segmentation Framework from 3D Late Gadolinium-Enhanced MRI using V-Net Family Models

Few-Shot Left Atrial Wall Segmentation in 3D LGE MRI via Meta-Learning