WE-E-213CD-11: A New Automatically Generated Metric for Evaluating the Spatial Precision of Deformable Image...

Area of Science:

• Medical imaging
• Image registration
• Quantitative analysis

Background:

• Image registration is crucial for comparing medical images across different time points or patients.
• Evaluating the accuracy of image registration, especially with inter-patient variability, remains a challenge.
• Existing methods often require ground truth or contoured structures, limiting their applicability.

Purpose of the Study:

• To introduce and validate a novel metric, Distance Discordance (DD), for quantifying image registration accuracy.
• To assess the performance of DD in evaluating the impact of anatomical variations on registration quality.
• To demonstrate the utility of DD as an interpretable and objective measure in medical image analysis.

Main Methods:

• Developed Distance Discordance (DD) as the distance between corresponding anatomical points in deformed images.
• Utilized a software phantom with varying object configurations and two B-Spline DIR algorithms (Elastix, Plastimatch) for phantom deformation.
• Calculated Distance Discordance Histograms (DDH) from voxel displacement distributions and applied the metric to Head & Neck patient data.

Main Results:

• Different image registration algorithms (Elastix, Plastimatch) yielded distinct DDH results.
• Mean DDH values were slightly lower for Elastix (0-1.28 cm) compared to Plastimatch (0-1.43 cm).
• DDH in Head & Neck patients followed a lognormal distribution (mean 0.45 cm, std. dev. 0.42 cm).

Conclusions:

• Distance Discordance (DD) provides an interpretable, quantitative method for assessing image registration goodness.
• DDH analysis can identify and potentially exclude images affected by significant inter-patient anatomical variability.
• The DD metric is advantageous as it does not rely on ground truth or contoured structures.