Multi-Scale Feature Fusion with Image-Driven Spatial Integration for Left Atrium Segmentation from Cardiac MR Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Disease Research

Background:

• Accurate left atrium (LA) segmentation in cardiac MRI is crucial for diagnosing cardiovascular diseases and planning atrial fibrillation (AF) ablation therapy.
• Manual segmentation is time-consuming and suffers from inter-observer variability, necessitating automated solutions.
• Class-agnostic foundation models offer feature extraction but may lack medical domain specificity, potentially reducing spatial resolution for fine anatomical details.

Purpose of the Study:

• To develop and validate an automated segmentation framework for left atrium (LA) in cardiac MRI.
• To enhance segmentation accuracy by integrating a foundation model (DINOv2) with a UNet-style decoder and multi-scale feature fusion.
• To address the limitations of reduced spatial resolution in foundation models for medical imaging tasks.

Main Methods:

• Proposed a segmentation framework combining DINOv2 as an encoder with a UNet-style decoder.
• Incorporated multi-scale feature fusion and input image reintroduction during decoding to preserve high-resolution details.
• Implemented a learnable weighting mechanism to dynamically prioritize hierarchical features from DINOv2 encoder blocks.

Main Results:

• Achieved a Dice score of 92.3% and an IoU score of 84.1% for the giant architecture on the LAScarQS 2022 dataset.
• Demonstrated superior performance compared to the nnUNet baseline model.
• Validated the framework's efficacy in improving automated left atrium segmentation from cardiac MRI.

Conclusions:

• The proposed framework effectively enhances automated left atrium segmentation accuracy in cardiac MRI.
• Integration of foundation models with domain-specific adaptations shows significant promise for medical image analysis.
• This approach advances the diagnosis and management of cardiovascular diseases, particularly AF, through improved imaging analysis.