Automated Ultrasound Uniformity: Correlating a Hybrid Quantitative Metric with Expert Assessment
Abstract
Purpose
Visual assessment of ultrasound uniformity is subjective and prone to inter-observer variability. This study validates a novel, automated metric combining weighted intensity statistics and texture analysis to provide an objective, continuous scale for uniformity, correlated with expert visual ranking.
Methods
The hybrid metric integrates: (1) a Spatially Weighted Standard Deviation (SWSD) using a sigmoid function to prioritize darker regions where non-uniformities are clinically critical, and (2) Local Binary Pattern (LBP) analysis to quantify local texture consistency. Images are processed in normalized patches, where the final index is calculated as the relative difference between the best and worst patches. The contributions of the two components are weighted by α and (1-α). Previous work established the algorithm’s sensitivity to simulated degradation. This study utilizes clinically acquired quality control (QC) images. Uniformity phantom images were consistently acquired by the same sonographer from 6 machines and 45 transducers, spanning 5 years at a breast imaging center. Eight images for visual assessment were selected across the range of automated uniformity scores. An ultrasound physicist visually scored the images on a scale of 1 (unacceptable) to 5 (excellent). Analysis was then performed to calculate correlation between the visual score and the algorithm’s uniformity index.
Results
The hybrid uniformity index closely tracked the physicist’s visual scoring and improved with increased SWHM weighting. The strongest agreement occurred at α = 0.75, yielding Spearman ρ = 0.945 (p = 0.0004) and Kendall τ = 0.866 (p = 0.005), indicating a robust monotonic relationship between automated uniformity and expert assessment. Lower α values (greater LBP influence) reduced concordance.
Conclusion
A hybrid SWHM–LBP metric provides an objective ultrasound uniformity score that closely matches expert visual grading, with α = 0.75 identified as the optimal weighting for this dataset.