Probabilistic inference of regularisation in non-rigid registration.

Area of Science:

• Medical image analysis
• Computational anatomy
• Biomedical imaging

Background:

• Non-rigid image registration requires careful selection of regularization to balance smoothness and complexity.
• Existing methods often use fixed, hand-tuned regularization levels, which are suboptimal for varying data quality and image pairs.
• Optimal regularization depends on image noise and anatomical similarity, necessitating adaptive strategies.

Purpose of the Study:

• To develop a probabilistic framework for inferring the regularization level directly from image data in non-rigid registration.
• To provide estimates of registration uncertainty as a byproduct of the probabilistic approach.
• To demonstrate the framework's effectiveness in inter-subject brain registration.

Main Methods:

• A probabilistic registration framework was developed to infer regularization from data.
• The framework was implemented using a free-form deformation transformation model.
• The method was applied to inter-subject brain registration of healthy control subjects.

Main Results:

• The framework successfully adapted regularization levels in the presence of image noise.
• Inferring regularization individually reduced model over-fitting, evidenced by decreased image folding.
• The adaptive approach achieved comparable overlap metrics to fixed methods while mitigating over-fitting.

Conclusions:

• The proposed probabilistic framework offers an adaptive solution for determining regularization in non-rigid image registration.
• This data-driven approach enhances registration robustness, particularly in noisy conditions.
• The framework provides valuable uncertainty estimates and reduces over-fitting, improving registration quality.