S-Net: a multiple cross aggregation convolutional architecture for automatic segmentation of small/thin structures...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Fully convolutional networks (FCNs) like U-Net are successful in medical image segmentation.
• However, deep FCNs struggle with detecting small/thin structures due to increasing receptive fields.
• This limits their effectiveness for structures like atrial walls and small arteries.

Purpose of the Study:

• To develop an improved FCN architecture for accurate segmentation of small and thin medical image structures.
• To address the limitations of traditional FCNs in capturing fine-grained details in deeper layers.

Main Methods:

• Proposed a novel S-shaped multiple cross-aggregation segmentation architecture (S-Net).
• S-Net utilizes two encoding branches: a resampling branch for low-level details and a downsampling branch for high-level knowledge.
• Employs residual cross-aggregation and lateral connections for feature fusion and supervised prediction at all decoding layers.

Main Results:

• S-Net demonstrated superior performance in segmenting cardiac walls and intracranial aneurysm (IA) vasculature.
• Quantitative and qualitative evaluations confirmed its effectiveness in predicting small/thin structures.
• The architecture successfully constrained receptive field growth, preserving fine details.

Conclusions:

• S-Net offers a significant advancement in medical image segmentation, particularly for challenging small/thin structures.
• The proposed architecture effectively combines low-level details and high-level semantics for improved segmentation accuracy.
• This method holds promise for enhanced diagnostic capabilities in various medical imaging applications.