Statistical biases correction in channelized Hotelling model observers.

Abstract

Objective: Channelized Hotelling model observers are efficient at simulating the human observer visual performance in medical imaging detection tasks. However, channelized Hotelling observers (CHO) are subject to statistical biases from zero-signal and finite-sample effects. The point estimate of the d' value is also not always symmetric with exact confidence interval (CI) bounds determined for the infinitely trained CHO. A method for correcting these statistical biases and CI asymmetry is studied.

Approach: CHO d' values and CI bounds with hold-out and resubstitution methods were computed for a range of 200x200 pixels images from 20 to 10 000 images for 10, 40 and 96 channels from a set of 20 000 images with gaussian coloured simulated noise and simulated signal. The median of the non-central F cumulative distribution (F'), which is the CHO underlying statistical behaviour for the resubstitution method, was computed, and compared to d' values and CI bounds. A set of experimental data was used to evaluate F' median values.

Main results: The F' median allows to get accurate corrected simulated d' values down to zero-signals. For small d' values, the variation of d' values with the inverse of number of images is not linear while the F' median allows a good correction in such conditions. The F' median is also inherently symmetric with regards to the confidence interval. With experimental data, F' median values in a range of about 1 to 10 d' values were within -0.8% to 4.7% of linearly extrapolated values at an infinite number of images.

Significance: The F' median correction is an effective simultaneous correction of the zero-signal statistical bias and finite-sample statistical bias, and of confidence interval asymmetry of channelized Hotelling observers. .