3D Isotropic High-Resolution Fetal Brain MRI Reconstruction From Motion Corrupted Thick Data Based on...

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• High-quality 3D fetal brain MRI is vital for diagnosing developmental abnormalities and understanding brain growth.
• Current methods for reconstructing 3D fetal brain MRI from 2D slices face challenges with motion artifacts and require extensive 3D training data.
• Deep learning (DL) shows promise for improving slice-to-volume registration (SVR) and super-resolution reconstruction (SRR), but clinical data limitations hinder most DL approaches.

Purpose of the Study:

• To develop an unsupervised, iterative, joint deep learning framework for 3D isotropic high-resolution (HR) fetal brain MRI volume reconstruction.
• To overcome the reliance on large-scale external 3D HR training datasets, which are difficult to obtain in clinical fetal MRI settings.
• To enhance the precision of fetal brain MRI analysis through improved reconstruction quality.

Main Methods:

• Proposed an unsupervised iterative joint SVR and SRR deep learning framework for 3D isotropic HR volume reconstruction.
• Conceptualized SVR as a convolutional neural network (CNN) that predicts rigid transformation matrices to align 2D slices with a 3D target volume.
• Employed a decoding network within a deep image prior framework for SRR, guided by local consistency and a comprehensive image degradation model.

Main Results:

• The proposed unsupervised DL framework successfully reconstructs high-quality 3D fetal brain MRI volumes from motion-corrupted 2D slices.
• Demonstrated superior performance compared to existing state-of-the-art fetal brain MRI reconstruction methods on both simulated and clinical datasets.
• Validated the effectiveness of the deep image prior framework in guiding HR volume reconstruction and the CNN in accurate slice-to-volume registration.

Conclusions:

• The developed unsupervised iterative joint SVR and SRR DL framework offers a robust solution for 3D fetal brain MRI reconstruction without external 3D training data.
• This method significantly advances the potential for precise clinical diagnosis and research into fetal brain development using MRI.
• The framework's ability to handle motion corruption and improve resolution holds promise for broader applications in medical imaging.