FCN Based Label Correction for Multi-Atlas Guided Organ Segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Anatomy

Background:

• Atlas-based segmentation is crucial for medical image analysis, offering robustness against subject variability.
• Accurate image registration is vital for atlas-based methods, but anatomical variations introduce inevitable errors.
• Existing methods struggle to mitigate the impact of registration inaccuracies on final segmentation performance.

Purpose of the Study:

• To develop a novel deep learning-based confidence estimation method to address registration errors in multi-atlas segmentation.
• To improve the robustness and accuracy of medical image segmentation by correcting for registration inaccuracies.
• To enhance the reliability of automated segmentation of anatomical structures.

Main Methods:

• A fully convolutional network (FCN) with residual connections was proposed to learn label confidence from image patch pairs.
• The FCN generates a label confidence map to identify and correct potential errors in warped atlas labels.
• Two label fusion techniques were employed to integrate the corrected atlas labels for final segmentation.

Main Results:

• The proposed deep learning approach effectively estimates label confidence, enabling the identification of registration errors.
• Correction of warped atlas labels based on confidence estimation led to improved segmentation accuracy.
• Validated on hippocampus segmentation, the method demonstrated superior performance compared to state-of-the-art techniques.

Conclusions:

• Deep learning-based confidence estimation offers a powerful strategy to mitigate registration errors in multi-atlas medical image segmentation.
• The proposed method enhances the accuracy and reliability of automated segmentation pipelines.
• This approach holds significant potential for improving clinical applications requiring precise anatomical segmentation.