Researchers have calculated the ‘segmentation ceiling’: masks accurate to within 10 per cent would be required to outperform direct regression on video, whereas a typical segmenter is off by 14 per cent.
The left ventricular ejection fraction, measured by echocardiography, guides many decisions in cardiology, and deep learning can estimate it directly from the video. Since it is clinically calculated from ventricular volumes, it is often assumed that adding explicit segmentation of the ventricle would improve the prediction. A study available on arXiv makes this intuition testable with a criterion called the "segmentation ceiling".
The mask must be very precise to serve
Starting from the ejection fraction formula, the authors derive analytically how the surface error of a mask, frame by frame, propagates to the final estimate. On the EchoNet-Dynamic game, the break-even point is around 10% area error per frame, while a representative segmenter is around 14%. Logically, four ways of injecting segmentation information fail to beat a model fed by raw video; ground truth masks only help because they betray the answer.
The real lever would be generalization: averaging the weights and a strong increase in the data give an R² of 0.806 (mean absolute error of 4.08 points), at the level of an R(2+1)D reference. Finally, a probabilistic formulation provides a well-calibrated uncertainty for each prediction, higher on the low ejection fractions, clinically the most delicate.
Read the full article on the original source
Go to the source →