Generative Adversarial Networks for Brain MRI Synthesis: Impact of Training Set Size on Clinical Application.

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Generative adversarial networks (GANs) are increasingly used for medical image synthesis.
• Evaluating the impact of training data size is crucial for optimizing GAN performance in clinical applications.

Purpose of the Study:

• To assess how training set size affects the performance of GANs in synthesizing brain MRI sequences (pre-contrast T1 and FLAIR).
• To compare GAN-generated MRIs with original images and evaluate their utility in a downstream segmentation task.

Main Methods:

• Three sets of GANs were trained: baseline (135 cases), early checkpoint (1251 cases), and late checkpoint (1251 cases).
• Models generated pre-contrast T1 (gT1) from post-contrast T1 and FLAIR (gFLAIR) from T2.
• Performance was evaluated using Structural Similarity Index (SSI) and Mean Squared Error (MSE) on a test set of 485 gliomas.
• Synthesized MRIs were used as input for a segmentation model, with results compared to original segmentations using Dice Similarity Coefficient (DSC).

Main Results:

• GANs trained on a smaller dataset (135 cases) performed comparably to those trained on a significantly larger dataset (1251 cases).
• For gT1, median SSI ranged from .918 to .957, and median MSE ranged from .006 to .014.
• For gFLAIR, median SSI ranged from .908 to .924, and median MSE ranged from .016 to .019.
• Segmentation performance (DSC) showed a wide range (.420-.955) but was generally comparable across different training set sizes.

Conclusions:

• Training set size has a limited impact on the performance of GANs for brain MRI synthesis.
• GANs are a practical and effective tool for generating synthetic brain MRI sequences, even with limited training data.
• This approach is particularly valuable for rare diseases and institutions with restricted data resources.