Self-supervised multicontrast super-resolution for diffusion-weighted prostate MRI.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Diffusion-weighted imaging (DWI) is crucial for cancer detection but often suffers from low resolution and signal-to-noise ratio (SNR).
• Conventional methods to boost SNR at high b-values compromise image resolution due to motion artifacts.
• Enhancing spatial resolution in DWI is essential for improved diagnostic accuracy.

Purpose of the Study:

• To develop a novel method for enhancing spatial resolution and SNR in diffusion-weighted images (DWI).
• To overcome the limitations of traditional DWI acquisition and processing techniques.
• To improve the diagnostic quality of DWI for cancer detection.

Main Methods:

• A "Perturbation Network" was developed to learn subvoxel-scale motions between acquisitions.
• An implicit neural representation (INR) network was jointly trained to encode DWI data as a continuous function.
• The model predicts higher-resolution signal intensities by evaluating the INR on a finer grid, correcting for motion using the Perturbation Network.

Main Results:

• Super-resolution images demonstrated significantly higher structural similarity to ground truth high-resolution images compared to traditional interpolation methods (p < 0.005).
• Blind qualitative assessments showed superior diagnostic quality for super-resolution images over interpolated images.
• The method effectively corrected for millimetric motion between acquisitions.

Conclusions:

• High-resolution details in DWI can be achieved without requiring high-resolution training data.
• The proposed method is adaptable to various scanner settings and anatomical regions, unlike supervised approaches.
• This technique offers a significant advancement for clinical DWI analysis and cancer detection.