Learning deep similarity metric for 3D MR-TRUS image registration

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Transrectal ultrasound (TRUS) and magnetic resonance (MR) image fusion enhances prostate biopsy accuracy for aggressive cancers.
• Robust automatic registration between MR and TRUS images is challenging due to significant appearance differences.
• Effective image registration is crucial for successful MR-TRUS fusion in targeted prostate biopsies.

Purpose of the Study:

• To develop a robust automatic registration method for MR-TRUS image fusion.
• To address challenges in defining a suitable similarity metric and optimization strategy for MR-TRUS registration.
• To improve the accuracy and reliability of image registration in the context of prostate cancer diagnostics.

Main Methods:

• A deep convolutional neural network was employed to learn a similarity metric for MR-TRUS registration.
• A composite optimization strategy was utilized to explore the solution space and find optimal initialization for second-order optimization.
• A multi-pass approach was implemented to smooth the learned similarity metric for improved optimization.

Main Results:

• The learned similarity metric demonstrated superior performance compared to classical mutual information and MIND feature-based methods.
• The registration framework exhibited a large capture range, indicating robustness.
• The proposed deep similarity metric-based approach achieved a mean target registration error (TRE) of 3.86 mm, significantly improving upon an initial TRE of 16 mm.

Conclusions:

• Deep neural network-learned similarity metrics can effectively assess image registration quality.
• The proposed framework enables robust automatic MR-TRUS registration, even with poor initialization.
• This approach enhances the reliability of image fusion for targeted prostate biopsies, potentially improving cancer detection rates.