Generating Synthetic MR Spectroscopic Imaging Data with Generative Adversarial Networks to Train Machine Learning...

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Machine learning models require large datasets for training, which are often scarce for specialized medical imaging techniques like MRSI.
• Current methods for acquiring MRSI data can be time-consuming and may not be part of routine MRI examinations.

Purpose of the Study:

• To develop a novel method for generating synthetic MR spectroscopic imaging (MRSI) data.
• To enable the training of machine learning models using readily available MRI and SVS data.

Main Methods:

• A pix2pix generative adversarial network model was employed to synthesize MRSI data.
• The model utilized T1- and T2-weighted MRI and single voxel spectroscopy (SVS) data as inputs.
• Quantitative evaluation involved Mean Squared Error (MSE) analysis and metabolite ratio comparisons.

Main Results:

• Synthetic MRSI data visually approximated reference data.
• Metabolite ratio confidence intervals largely overlapped with reference values.
• Lower MSE was observed when SVS data was in the same location as reference MRSI.

Conclusions:

• A novel method effectively generates synthetic MRSI data by integrating MRI and SVS.
• This technique can augment MRSI datasets for machine learning, potentially improving model performance.
• Integrating SVS acquisition into routine MRI offers a pathway to increase training data volume.