Expert variability as a benchmark for validating automatic CT-MRI registration in brain stereotactic radiosurgery.

Area of Science:

• Medical Imaging
• Radiosurgery
• Computational Anatomy

Background:

• Accurate computed tomography (CT) and magnetic resonance imaging (MRI) registration is essential for stereotactic radiosurgery (SRS) of brain metastases (BM).
• Existing manual registration methods exhibit variability, necessitating a reliable benchmark for algorithm validation.
• A probabilistic gold standard (GS) and acceptance thresholds are needed to evaluate automatic CT-MRI registration algorithms.

Purpose of the Study:

• To establish a probabilistic gold standard (GS) for rigid CT-MRI registration using multi-expert variability.
• To define clinically meaningful acceptance thresholds for automatic registration algorithms.
• To evaluate the performance of mutual-information (MI)-based registration methods, including a contour-based (CB) approach.

Main Methods:

• Twenty CT-MRI pairs with 39 brain metastases (BM) were registered by six experts twice.
• Variability was quantified using bootstrap resampling to estimate confidence intervals for key registration metrics.
• A contour-based (CB) mutual-information (MI) registration algorithm, using manual or AI-generated contours, was evaluated against expert performance.

Main Results:

• Expert CT-MRI registration showed low but significant variability (99% CI: 0.31° rotation, 0.34 mm translation).
• The contour-based (CB) algorithm achieved performance within expert acceptance ranges for most metrics, irrespective of contour source (manual vs. AI).
• Less-experienced operators' deviations were reduced when using CB-aligned datasets, highlighting the algorithm's utility.

Conclusions:

• Quantifying multi-expert variability provides a probabilistic gold standard and relevant thresholds for CT-MRI registration validation.
• The AI-driven CB algorithm demonstrates high reliability and reduces operator dependence in CT-MRI registration.
• This validated benchmark and reliable algorithm can streamline workflows for brain metastases SRS.