Semi-Supervised Learning of MRI Synthesis Without Fully-Sampled Ground Truths.

Area of Science:

• Medical Imaging
• Deep Learning
• Artificial Intelligence

Background:

• Supervised deep learning for MRI contrast translation requires fully-sampled data, which is often impractical due to scan time and cost limitations.
• Existing methods struggle with undersampled data, limiting the feasibility of learning-based multi-contrast MRI synthesis.

Purpose of the Study:

• To introduce the first semi-supervised model for MRI contrast translation (ssGAN) capable of training directly on undersampled k-space data.
• To develop novel multi-coil losses for enabling semi-supervised learning on undersampled MRI data.

Main Methods:

• The ssGAN model utilizes novel multi-coil losses applied selectively to acquired k-space samples across image, k-space, and adversarial domains.
• Experiments were conducted using retrospectively undersampled multi-contrast brain MRI datasets.

Main Results:

• The ssGAN model achieved performance comparable to supervised models trained on fully-sampled data.
• ssGAN outperformed single-coil models and cascaded reconstruction-synthesis models trained on undersampled data.
• The proposed multi-coil losses effectively leverage acquired k-space samples for improved synthesis.

Conclusions:

• The ssGAN model significantly improves the feasibility of learning-based multi-contrast MRI synthesis by enabling training with undersampled data.
• This approach reduces the need for high-quality, fully-sampled datasets, potentially lowering MRI acquisition costs and time.
• ssGAN offers a promising solution for efficient and effective MRI contrast translation in clinical settings.