Learning Geodesic Active Contours for Embedding Object Global Information in Segmentation CNNs.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Current Convolutional Neural Networks (CNNs) for image segmentation primarily use local image features.
• This reliance on local information often neglects crucial object global context, leading to segmentation inaccuracies, particularly at boundaries.

Purpose of the Study:

• To develop an end-to-end deep learning framework that incorporates global geometric information for more accurate image segmentation.
• To leverage classical geodesic active contours (GAC) within a novel level set function (LSF) regression network to improve segmentation performance.

Main Methods:

• A level set function (LSF) regression network is proposed, trained using segmentation and LSF ground truth.
• The network is supervised to directly minimize the geodesic active contours (GAC) energy functional.
• This approach embeds global geometric constraints into the segmentation process.

Main Results:

• The proposed method significantly outperforms state-of-the-art (SOTA) learning active contour and boundary error reduction methods.
• Experiments demonstrate superior or competitive results on multi-class segmentation tasks, including organ and tumor segmentation.
• The integration of GAC effectively reduces boundary errors and segmentation outliers.

Conclusions:

• Incorporating global information via GAC substantially improves image segmentation accuracy, especially in reducing boundary errors and outliers.
• The method shows significant potential for applications requiring high segmentation precision, such as surgical planning and image registration.