Machine Learning for Medical Image Translation: A Systematic Review.

Area of Science:

• Medical imaging
• Artificial intelligence
• Radiology

Background:

• Computed tomography (CT) scans are cost-effective for initial neurological assessments.
• Magnetic resonance imaging (MRI) offers superior detail for detecting abnormalities.
• Deep learning (DL) is increasingly used for medical image synthesis.

Purpose of the Study:

• To review studies utilizing deep learning for synthetic medical image generation.
• To analyze trends in MRI to CT and CT to MRI synthesis.
• To identify motivations and limitations in medical image synthesis research.

Main Methods:

• A comprehensive literature search was conducted in March 2023.
• 103 relevant studies published since 2017 were selected and analyzed.
• Analysis included publication year, dataset size, modalities, DL architecture, motivations, and evaluation methods.

Main Results:

• 74% of studies focused on MRI to CT synthesis, often for MRI-only radiation therapy.
• Other synthesis directions included CT to MRI, Cross MRI, PET to CT, and MRI to PET.
• Key motivations were supporting MRI-only radiotherapy, aiding diagnosis, and augmenting datasets.

Conclusions:

• MRI to CT synthesis is more researched than CT to MRI, despite the latter's benefits.
• Limited availability and size of medical datasets, especially paired ones, pose a significant challenge.
• Recommendations include forming a global consortium for dataset acquisition and establishing standardized evaluation methods for synthesized images in clinical practice.