Post-DAE: Anatomically Plausible Segmentation via Post-Processing With Denoising Autoencoders.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Biomedical image segmentation often requires post-processing to ensure anatomical plausibility.
• Existing methods struggle to incorporate complex priors like topological restrictions or convexity.
• Denoising autoencoders (DAE) offer advanced manifold learning capabilities.

Purpose of the Study:

• To introduce Post-DAE, a novel post-processing technique for improving biomedical image segmentation.
• To leverage DAE for learning a manifold of anatomically plausible segmentations.
• To reconstruct segmentation masks onto this learned manifold for enhanced accuracy.

Main Methods:

• Utilizing denoising autoencoders (DAE) for manifold learning.
• Training Post-DAE with unpaired segmentation masks, independent of image modality or intensity.
• Reconstructing arbitrary segmentation masks by projecting them onto the learned manifold.

Main Results:

• Demonstrated improvement in binary and multi-label segmentation of chest X-ray and cardiac MRI.
• Successfully corrected erroneous and noisy segmentation masks.
• Achieved enhanced segmentation plausibility with negligible additional computation.

Conclusions:

• Post-DAE effectively refines biomedical image segmentation masks.
• The method enhances anatomical plausibility by learning and projecting onto a manifold of valid segmentations.
• Post-DAE offers a computationally efficient solution for improving segmentation accuracy across various modalities.