Three-dimensional self super-resolution for pelvic floor MRI using a convolutional neural network with...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• High-resolution pelvic MRI is crucial for evaluating pelvic floor disorders (PFDs).
• Current 3D MRI acquisition is time-consuming, leading to low through-plane resolution.
• Limited 3D data necessitates improved resolution techniques for accurate PFD assessment.

Purpose of the Study:

• To develop a 3D super-resolution technique for pelvic MRI using convolutional neural networks (CNNs).
• To enhance through-plane resolution by leveraging intrinsic similarities in low-resolution 3D MR data from multiple orientations.
• To improve the evaluation of pelvic floor disorders through advanced MRI post-processing.

Main Methods:

• A 2D super-resolution CNN model, specifically the residual-in-residual dense block network (RRDBNet), was employed.
• The RRDBNet was sequentially applied to projected low-resolution views (axial, sagittal, coronal) to achieve super-resolution.
• Experiments utilized three datasets (two private, one public) to train, test, and validate the model, comparing it against interpolation and EDSR methods.

Main Results:

• A CNN-based method successfully learned similarities among MR acquisitions from different planes.
• Achieved through-plane super-resolution for pelvic MR images without requiring high-resolution 3D data.
• Demonstrated effectiveness in improving 3D geometric model reconstruction for tasks like urinary bladder analysis.

Conclusions:

• The developed CNN approach provides effective 3D super-resolution for pelvic MRI.
• This method enhances image quality and is beneficial for the analysis of pelvic floor disorders.
• The technique offers a valuable tool for improving diagnostic accuracy in pelvic MRI.