Exploiting Global Structure Information to Improve Medical Image Segmentation.

Area of Science:

• Medical imaging
• Computer vision
• Machine learning

Background:

• Accurate segmentation of anatomical structures in medical images is crucial for diagnosis and treatment planning.
• Existing methods often rely on pre-trained models or separate training stages, limiting performance and adaptability.

Purpose of the Study:

• To propose a novel single-stage method for enhancing medical image segmentation performance.
• To improve the learning of global anatomical structures and boundary information.
• To enhance the robustness of segmentation models against domain shifts.

Main Methods:

• A convolutional neural network-based approach is proposed, integrating an autoencoder for learning global boundary structures.
• The autoencoder and segmentation network are jointly learned in a single stage, constraining the segmentation through a loss function.
• The method enables segmentation predictions within a learned anatomical feature space.

Main Results:

• The proposed method significantly enhances segmentation performance on lung area and spinal cord datasets, as validated by overlap and distance metrics.
• Experimental results demonstrate improved robustness against domain shifts compared to previous approaches.
• The joint learning strategy effectively captures global anatomical context.

Conclusions:

• The proposed single-stage joint learning method offers a superior approach to medical image segmentation.
• This technique improves both segmentation accuracy and model robustness, paving the way for more reliable clinical applications.
• The method's ability to learn anatomical features directly contributes to enhanced performance.