Bias in Hotelling Observer Performance Computed from Finite Data.

Area of Science:

• Medical imaging analysis
• Signal detection theory
• Statistical modeling

Background:

• The Hotelling observer maximizes test-statistic SNR for signal detection in images.
• Calculating Hotelling SNR typically requires inverting large covariance matrices.
• Recent methods enable covariance matrix inversion with limited image data.

Purpose of the Study:

• To investigate the bias in Hotelling trace estimates using covariance matrix decomposition.
• To evaluate this bias in both high and low detector-noise scenarios.
• To present a theoretical analysis and simulation studies of Hotelling trace bias.

Main Methods:

• Utilized covariance matrix decomposition separating detector-noise and background variability.
• Applied methods for estimation and inversion of large covariance matrices with few images.
• Conducted theoretical evaluation and extensive simulation studies.

Main Results:

• Covariance matrix decomposition allows for the creation of a full-rank, invertible covariance matrix with limited data.
• The decomposition method is generally effective in high-detector-noise situations.
• Significant bias in Hotelling trace estimates was observed in extremely low-noise conditions.

Conclusions:

• Covariance matrix decomposition is a valuable tool for Hotelling observer analysis in medical imaging.
• The method's performance is dependent on noise levels, with potential bias in low-noise environments.
• Further theoretical and simulation-based research is needed to address bias in low-noise scenarios.