Landmark matching and B-spline implicit neural representations for diffusion-weighted imaging distortion correction.

Area of Science:

• Medical Imaging
• Radiotherapy
• Computational Anatomy

Background:

• Geometric distortions in diffusion-weighted imaging (DWI) impede precise tumor delineation and localization for radiation therapy.
• Conventional mutual information optimization methods for correcting these distortions often result in non-smooth and physically implausible deformations due to local minima.

Purpose of the Study:

• To propose and evaluate a novel Landmark Matching B-spline Implicit Neural Representation (LMBS-INR) framework for accurate DWI distortion correction.
• To overcome the limitations of traditional optimization methods by integrating anatomical landmark correspondences and B-spline deformation fields.

Main Methods:

• The LMBS-INR framework utilizes a foundation landmark matching model to establish anatomical correspondences.
• B-spline deformation fields, modeled using Fourier-encoded multi-layer perceptrons, ensure physically plausible transformations for robust multimodal registration between DWI and anatomical references.

Main Results:

• The proposed method demonstrated superior performance on brain and abdominal datasets, achieving high Dice coefficients (0.919 ± 0.038 for brain, 0.926 ± 0.032 for abdomen).
• Evaluations on simulated data yielded excellent metrics, including PSNR (25.912 ± 3.148 dB), NCC (0.911 ± 0.137), and SSIM (0.888 ± 0.107), outperforming all baseline methods.

Conclusions:

• The LMBS-INR framework effectively corrects geometric distortions in DWI by combining B-spline parameterization with foundation model capabilities.
• This approach offers enhanced precision for intra/post-radiotherapy assessment, improving the utility of DWI in clinical practice.