Reference-free Bayesian model for pointing errors of typein neurosurgical planning.

Area of Science:

• Neurosurgical imaging analysis
• Medical image segmentation
• Computational anatomy

Background:

• Identifying critical points in volumetric neurosurgical images is challenging due to inter-annotator variability.
• Inferred points, not directly visible, lead to significant disagreement among experts, complicating research.
• Errors of type, where experts select fundamentally different points, are a key challenge.

Purpose of the Study:

• To develop a regularized Bayesian model for measuring "errors of type" in image-based pointing tasks.
• To create a reference-free model that does not require ground truth data.
• To assess annotator consensus for improved reliability in neurosurgical planning.

Main Methods:

• A regularized Bayesian framework was developed to quantify disagreement in pointing tasks.
• The model operates without prior knowledge of the ground truth point.
• It leverages the consensus level among multiple annotators to estimate accuracy.

Main Results:

• The model achieved an estimated probability of selecting the correct point between 82.6% and 88.6%.
• Uncertainty in these estimates ranged from 2.8% to 4.0%, providing dataset strength insights.
• Results align with literature values where ground truth was known.

Conclusions:

• The reference-free Bayesian model effectively quantifies "errors of type" in pointing tasks.
• Enables clinical studies with fewer annotators and unknown ground truth.
• Facilitates better understanding of human error in neurosurgical planning.