Random forest regression for magnetic resonance image synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Magnetic Resonance Imaging (MRI) offers diverse tissue contrasts via flexible pulse sequences.
• Inconsistent MRI acquisitions across datasets/sessions can lead to unreliable automated image analysis.
• Existing image synthesis methods struggle with T2-weighted brain images (including skull) and FLAIR images.

Purpose of the Study:

• To introduce REPLICA, a novel supervised random forest approach for synthesizing MRI contrasts.
• To overcome limitations in synthesizing T2-weighted and FLAIR MRI images.
• To enable intensity standardization across different MRI datasets.

Main Methods:

• REPLICA employs a supervised random forest model for nonlinear regression.
• It learns to predict alternate tissue contrasts from input contrasts.
• The method synthesizes T2-weighted full head images and FLAIR images.

Main Results:

• REPLICA successfully synthesizes both T2-weighted and FLAIR MRI images.
• Experimental results validate synthetic images through direct comparison and image analysis tasks.
• REPLICA demonstrates computational efficiency and comparable performance to state-of-the-art methods.

Conclusions:

• REPLICA effectively addresses limitations in MRI image synthesis.
• The method enables consistent image analysis by standardizing intensity across datasets.
• REPLICA offers a versatile solution for generating T2-weighted and FLAIR MRI contrasts.