Robust automated preclinical fMRI preprocessing via a multi-stage dilated convolutional Swin Transformer affine...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomedical Engineering

Background:

• Accurate preprocessing of preclinical functional magnetic resonance imaging (fMRI) data is vital for reliable analysis.
• Traditional and deep learning methods face challenges due to low resolution, brain geometry variations, and limited datasets.

Purpose of the Study:

• To develop an advanced deep learning-based preprocessing pipeline for preclinical fMRI.
• To enhance affine registration accuracy for aligning fMRI data with standard atlases.

Main Methods:

• The pipeline integrates 3D Generative Adversarial Network (GAN)-based denoising and Transformer-based skull stripping.
• A novel Multi-stage Dilated Convolutional Swin Transformer (MsDCSwinT) is proposed for affine registration.
• The method captures both local and global spatial misalignments in preclinical fMRI data.

Main Results:

• The proposed pipeline was validated across multiple preclinical fMRI studies.
• The Swin Transformer-based affine registration module demonstrated superior performance.
• Higher average Dice similarity coefficients were achieved compared to state-of-the-art methods.

Conclusions:

• The developed pipeline provides a robust, accurate, and automated solution for preclinical fMRI preprocessing.
• Leveraging Generative Adversarial Networks and Transformers significantly improves data alignment and analysis.
• This work advances the capabilities for analyzing complex preclinical brain imaging data.