The Image-to-Physical Liver Registration Sparse Data Challenge: comparison of state-of-the-art using a common dataset

Area of Science:

• Medical Imaging
• Computational Anatomy
• Surgical Guidance

Background:

• Image-to-physical registration is crucial for surgical guidance.
• Sparse surface data and soft tissue deformation pose significant challenges to accurate anatomical alignment.
• The Image-to-Physical Liver Registration Sparse Data Challenge aimed to evaluate registration methods on a common dataset.

Purpose of the Study:

• To benchmark the performance of various image-to-physical registration algorithms using sparse liver surface data.
• To identify effective tactics and limitations in current registration methods.
• To inform the development of future image-to-physical registration algorithms.

Main Methods:

• Three rigid and five deformable registration algorithms were evaluated, including deep learning and biomechanical approaches.
• Algorithms were tested on a dataset of 112 registration scenarios using a tissue-mimicking phantom with 159 subsurface validation targets.
• Target Registration Errors (TRE) were assessed under varying data extents, target locations, and noise levels; Jacobian determinants and strain magnitudes were analyzed for displacement field consistency.

Main Results:

• Rigid registration methods showed significant variability in TRE.
• Two biomechanical methods achieved TRE of $$ and $$, outperforming optimal rigid registration.
• Biomechanical methods demonstrated robust performance across different surface data coverages and liver segments; deep learning methods showed TRE ranging from $$ to $$.

Conclusions:

• Algorithm choice critically influences registration accuracy and deformation field variability.
• Biomechanical simulations incorporating task-specific boundary conditions currently offer the best performance for sparse data-driven image-to-physical registration.
• Further development is needed for deep learning-based registration methods.