Optimizing Performance of Transformer-based Models for Fetal Brain MR Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Fetal brain extraction from resting-state functional MRI (rs-fMRI) is crucial for developmental studies.
• Transformer-based models offer potential for improving medical image segmentation tasks.

Purpose of the Study:

• To evaluate the performance of transformer models (Swin-UNETR) for fetal brain extraction.
• To investigate the impact of pretraining weights, dataset size, and input size on model performance.
• To compare the best-performing model against reference standards and state-of-the-art methods.

Main Methods:

• Utilized retrospective internal (172 fetuses) and external (131 fetuses) datasets.
• Investigated Swin-UNETR and UNETR models, manipulating dataset size, pretraining, and input image size.
• Assessed generalization across different scanner types and gestational weeks (GWs) using Dice Similarity Coefficient (DSC) and Balanced Average Hausdorff Distance (BAHD).

Main Results:

• Swin-UNETR demonstrated robustness to pretraining approach and dataset size, performing optimally with mean dataset image size (DSC: 0.92, BAHD: 0.097).
• The model showed consistent performance across different scanner types.
• Swin-UNETR achieved comparable performance to reference models in late-fetal periods (GWs > 25) but lower performance in mid-fetal periods (GWs ≤ 25).

Conclusions:

• Swin-UNETR exhibits flexibility with smaller datasets and various pretraining strategies for rs-fMRI fetal brain extraction.
• Transformer models show promise, particularly for late-fetal brain segmentation, though further refinement is needed for earlier stages.