Assessment of three methods for detection of ultrasound artifacts

The purpose of this work is to measure the performance (sensitivity and specificity) of three different visually based methods of detecting ultrasound artifacts using dynamic clips obtained with a liquid phantom. It is important to detect the presence of these non uniformities as early as possible, so they can be assessed, tracked, and addressed well before clinical image quality is impacted.

A total of 28 transducers of varying models containing a single artificial artifact created by stretching thin filaments across the transducer face were prepared. A second set of 28 matching transducers contained no artifacts. A 10 s clip was recorded of a dynamic speckle pattern from a custom liquid phantom ("dynamic clip"). A single-frame image was obtained by computing the median values at each pixel location over all frames of the clip ("median image"). This single-frame median image was then subtracted from a baseline image previously obtained with no induced artifact ("subtracted median"). All images were evaluated by six observers. The mean sensitivity and specificity with 95% exact binomial confidence intervals for the three artifact detection methods were estimated. Evaluation time and observer confidence were recorded.

Both single-frame median and subtracted median images had a higher sensitivity than the dynamic clip. Overall the subtracted median images had the highest sensitivity of 97%, while maintaining a high specificity of 92%. Observers identified artifacts in the shortest time and with the highest confidence with this method. The subtracted median method removed the original structural variations and non-uniformities, and reduced the likelihood of false-positives.

From the three methods assessed in this study, subtracted median images allow detection of artifacts with very good sensitivity and specificity, low image evaluation times, and a high degree of observer confidence, making these ideal for routine QC. For acceptance testing, where there are no previous baseline images available for subtraction, the use of median images is useful, although comparison with median images from different transducers of the same model and ∕ or comparison of impressions from multiple observers should be made to decrease the incidence of false-positive findings. If median and subtracted median images are not available, direct inspection of the dynamic B-mode clips is useful for acceptance testing and quality control, but with lower sensitivity and somewhat longer evaluation time.