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

2026-04-30 · Source: cs.CV updates on arXiv.org · Field: Science & Research — Health & Medical Research, Engineering & Applied Sciences, Mathematics & Computational Sciences · Depth: Advanced, medium

Summary

A multi-stage framework has been developed for multi-class bi-atrial segmentation from 3D late gadolinium-enhanced (LGE) MRI of the human heart, addressing a challenge held as part of MICCAI 2024. The pipeline incorporates a preprocessing step using multidimensional contrast limited adaptive histogram equalization (MCLAHE) to enhance raw MRI scans, followed by a two-stage segmentation process. The first stage performs coarse region detection of the atrium using a V-Net family model on down-sampled data, while the second stage executes fine multi-class segmentation (left atrium, right atrium, wall, and background) on a cropped region using another V-Net model. Asymmetric loss is applied for model optimization, with fixed hyperparameters γ₊=1, γ₋=4, and m=0.05. Experiments on a public validation set of 30 3D LGE-MRI scans showed that a vanilla V-Net model with MCLAHE preprocessing achieved the best performance, outperforming a more complex V-Net++ model in terms of Dice and HD95 metrics.

Key takeaway

For Computer Vision Engineers developing medical image segmentation models, you should prioritize robust preprocessing techniques like MCLAHE, as it significantly improves model performance and transferability, even with simpler architectures. Do not solely rely on training loss curves; validate models on diverse datasets to assess their real-world effectiveness, especially for imbalanced tasks like bi-atrial segmentation.

Key insights

A multi-stage V-Net framework with MCLAHE preprocessing effectively segments bi-atrial regions from 3D LGE-MRI.

Principles

• MCLAHE significantly enhances medical image segmentation.
• Simpler models can outperform complex ones in transferability.
• Multi-stage segmentation improves accuracy for small regions.

Method

The method involves MCLAHE preprocessing, coarse atrium region detection using a V-Net on down-sampled data, and fine multi-class segmentation on a cropped region with another V-Net, optimized by asymmetric loss.

In practice

• Apply MCLAHE for medical image enhancement.
• Consider multi-stage segmentation for small ROIs.
• Evaluate model transferability beyond training loss.

Topics

• Bi-atrial Segmentation
• 3D LGE-MRI
• V-Net Family Models
• MCLAHE Preprocessing
• Asymmetric Loss

Code references

• wenh06/MBAS2024

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Multi-layer feature aggregation network with residual module and attention mechanism for jaw cyst image segmentation — MFA-Net enhances medical image segmentation by integrating residual modules, channel-spatial attention, and multi-layer feature aggregation.
• ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements — ROBUST-WT demonstrates that targeted training enhancements can substantially improve medical image segmentation robustness without altering core architecture.
• FoR-Net: Learning to Focus on Hard Regions for Efficient Semantic Segmentation — FoR-Net uses region-focused reasoning and a Top-K activation to efficiently segment hard regions.
• ConvNeXt-FD: A Fractal-Based Deep Model for Robust Biomedical Image Segmentation — ConvNeXt-FD enhances biomedical image segmentation by combining ConvNeXt with fractal-inspired boundary regularization for superior shape fidelity.
• Transformer-Guided Graph Attention for Direct Cardiac Mesh Reconstruction: A Structural Digital Twin Framework — An end-to-end network directly reconstructs cardiac surface meshes from 3D medical images, bypassing traditional multi-step processes.
• MVSegNet: A Lightweight Boundary-Aware Network for Fetal Lateral Ventricle Segmentation and Atrial Width Estimation in Prenatal Ultrasound — MVSegNet accurately segments fetal lateral ventricles for atrial width estimation despite ultrasound challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.