Test-time adaptable neural networks for robust medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel in supervised learning but degrade with variations in medical image acquisition.
• Mismatches in scanner models or protocols between training and testing datasets cause significant performance drops in CNN-based segmentation.

Purpose of the Study:

• To develop a robust medical image segmentation method that maintains performance despite variations in imaging acquisition details.
• To enhance the generalizability of CNNs for segmentation tasks across different scanner models and protocols.

Main Methods:

• Designed a segmentation CNN comprising two sub-networks: an image normalization CNN and a deep segmentation CNN.
• Implemented test-time adaptation by fine-tuning the normalization sub-network for each test image.
• Utilized a denoising autoencoder (DAE) to model priors on plausible anatomical segmentation labels for guiding adaptation.

Main Results:

• Validated the approach on multi-center MRI datasets for brain, heart, and prostate imaging.
• Demonstrated consistent performance improvements through test-time adaptation.
• The proposed method proved effective in increasing robustness to variations in imaging scanners and protocols.

Conclusions:

• The proposed test-time adaptable network design offers a generalizable solution to improve CNN robustness in medical image segmentation.
• This architecture is compatible with various deep CNN segmentation networks, enhancing their applicability across diverse imaging conditions.