Unsupervised MRI Homogenization: Application to Pediatric Anterior Visual Pathway Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning excels in medical image analysis but requires large datasets, which are often unavailable for rare diseases and pediatric imaging.
• Magnetic Resonance Imaging (MRI) data variability is high due to non-standardized protocols and sedation needs in children.
• Existing methods struggle with diverse and limited medical imaging datasets.

Purpose of the Study:

• To develop a deep learning architecture for Magnetic Resonance Imaging (MRI) homogenization and simultaneous anatomical segmentation.
• To address data scarcity and variability challenges in pediatric and rare disease imaging.
• To create a unified representation of MRI data adaptable to various imaging protocols.

Main Methods:

• An unsupervised deep learning architecture combining variational autoencoder with cycle generative adversarial networks for MRI homogenization.
• Unpaired image-to-image translation to learn a common data space representing an optimal imaging protocol.
• Simultaneous supervised learning for generating segmentation maps of anatomical regions of interest.

Main Results:

• The proposed method successfully homogenized MRI datasets with variable protocols and vendors.
• Simultaneous segmentation of the anterior visual pathway was achieved.
• The homogenization technique significantly improved segmentation performance compared to a non-homogenized multi-protocol U-Net.

Conclusions:

• The presented deep learning architecture effectively homogenizes MRI data and performs segmentation simultaneously.
• This approach offers a robust solution for analyzing medical images from diverse and limited datasets, particularly in pediatric and rare disease contexts.
• The method enhances the reliability and accuracy of medical image analysis by overcoming protocol variability.