Robust and efficient abdominal CT segmentation using shape constrained multi-scale attention network.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Deep learning models for abdominal CT segmentation face challenges due to diverse image characteristics and organ variability.
• Robust segmentation is crucial for accurate diagnosis and treatment planning in various diseases.

Purpose of the Study:

• To develop a robust and efficient two-stage deep learning method for multi-organ segmentation in abdominal CT images.
• To address challenges posed by varying intensity distributions, organ shapes, and disease conditions.

Main Methods:

• A two-stage approach combining a binary segmentation network for coarse localization and a multi-scale attention network for fine segmentation.
• An auxiliary network pre-trained for shape feature learning was used to constrain the fine segmentation network, especially for organs affected by severe diseases.

Main Results:

• The method achieved an average Dice Similarity Coefficient (DSC) of 83.7% and Normalized Surface Dice (NSD) of 64.4% on the FLARE challenge dataset.
• The proposed approach secured second place among over 90 teams in the FLARE challenge (MICCAI 2021).

Conclusions:

• The developed method demonstrates promising robustness and efficiency for abdominal multi-organ segmentation.
• The findings suggest potential for advancing the clinical application of automated abdominal CT segmentation.