Comparison of auto-contouring and hand-contouring of ultrasound images of the tongue surface.

The study used only four speakers and 40 short clips, so it does not show how the methods perform on many clients or long recordings. The clips were about 748 ms long, much shorter than many real clinical or research recordings, so the authors could not tell whether some errors would grow, plateau, or correct with more time. The tolerance thresholds were based on human measurer differences plus an arbitrary margin, so passing or failing a threshold is not a fixed clinical accuracy standard. The hand contours were used as the standard, but there was no independent ground truth such as a calibrated concurrent imaging signal. All data came from one ultrasound system with fixed settings, and harder-to-see images may have made differences larger. The AAA methods were run on video files rather than original Cine Loop data, so their performance on Cine Loop data may be different. SLURP output changed slightly each run, although the authors judged that this did not meaningfully affect results.

The article text says the authors report no conflicts of interest. The supplied metadata indicates NIH extramural support, but no sponsor role is stated.

Do not conclude that automatic tongue contouring is accurate enough for all speech sounds without checking. The study found that no algorithm met all thresholds, and all algorithms were much less reliable at the anterior tongue-tip end than human measurers.