Double-branch U-Net for multi-scale organ segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• U-Net is a successful convolutional neural network architecture for medical image segmentation.
• Standard U-Net architectures can struggle with multi-scale segmentation tasks, particularly with small objects, due to excessive downsampling.
• Accurate segmentation of multi-scale medical images is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To propose a novel deep learning architecture, Double-branch U-Net (2BUNet), to overcome limitations of existing models in multi-scale medical image segmentation.
• To enhance the segmentation of small objects often missed by conventional U-Net models.
• To achieve state-of-the-art performance in real-world medical segmentation scenarios.

Main Methods:

• Introduced a Double-branch U-Net (2BUNet) comprising a main branch, a tributary branch, an information exchange module, and a classification module.
• The main branch focuses on extracting features from larger organs, while the tributary branch enlarges the image to capture microscopic features of smaller organs.
• An information exchange module provides regularization between branches, and a classification module adds class constraints to the output tensor.

Main Results:

• The proposed 2BUNet effectively addresses the challenge of ignoring small segmentation objects caused by excessive downsampling in standard U-Net models.
• Comparative experiments demonstrate that 2BUNet achieves superior segmentation performance compared to existing methods on multi-scale medical images.
• The model exhibits state-of-the-art performance in real-world medical segmentation applications.

Conclusions:

• The Double-branch U-Net (2BUNet) architecture significantly improves multi-scale medical image segmentation by effectively handling objects of varying sizes.
• The parallel branch design and information exchange mechanism enhance feature extraction and model robustness.
• 2BUNet represents a promising advancement for accurate and reliable medical image segmentation.