Multi-view hybrid encoder U-Net for 3D renal vascular medical image segmentation

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Anatomy

Background:

• Accurate segmentation of renal artery vessels is crucial for understanding cardiovascular health.
• Existing methods face challenges in capturing detailed morphological features like size, shape, and branching angles.

Purpose of the Study:

• To develop and evaluate a multi-view hybrid encoder U-Net for enhanced segmentation of renal artery vessels.
• To improve the extraction and representation of intricate vascular structures in human tissue.

Main Methods:

• A multi-view hybrid encoder U-Net architecture was proposed, utilizing a hierarchical encoder with lightweight CNN and hybrid CNN/Transformer modules.
• The model processes 2D images from three orthogonal views, preserving spatial resolution to maintain fine details.
• The approach focuses on enhancing feature extraction and representation for detailed segmentation.

Main Results:

• The proposed method achieved a Mean Surface Dice (MSD) value of 0.852 and a Dice Similarity Coefficient (DSC) of 0.939 on unseen kidney data.
• These results demonstrate high accuracy in segmenting renal artery vessels, preserving subtle morphological features.
• The hybrid encoding mechanism proved effective in enhancing feature extraction and representation.

Conclusions:

• The developed multi-view hybrid encoder U-Net is effective for detailed renal artery segmentation.
• The model's ability to preserve spatial resolution and capture fine details contributes to its high performance.
• Further validation across diverse datasets and modalities is recommended to confirm generalizability.