Statistical biases correction in channelized Hotelling model observers

Area of Science:

• Medical Imaging
• Observer Performance Modeling

Background:

• Channelized Hotelling observers (CHO) simulate human visual performance in medical imaging.
• CHO are susceptible to statistical biases from zero-signal and finite-sample effects.
• Point estimates of d' values and confidence intervals (CI) can be asymmetric.

Purpose of the Study:

• To study a method for correcting statistical biases and CI asymmetry in CHO.
• To evaluate the effectiveness of the F-prime median for bias correction.

Main Methods:

• Computed CHO d' values and CI bounds using hold-out and resubstitution methods across varying image numbers and channels.
• Calculated the median of the non-central F cumulative distribution (F') for the resubstitution method.
• Compared F' median values to d' values and CI bounds using simulated and experimental data.

Main Results:

• The F' median accurately corrects simulated d' values, even with zero signals.
• It provides good correction for small d' values where d' variation is non-linear with image count.
• The F' median inherently offers symmetric CI bounds.

Conclusions:

• The F' median effectively corrects zero-signal and finite-sample statistical biases in CHO.
• This method also addresses CI asymmetry, enhancing observer performance modeling in medical imaging.