WE-DE-207B-07: Hotelling Templates Without Additional Training Data for Breast Tomosynthesis Observer Models

Authors


Abstract

Purpose:

To describe and test a practical approach that uses minimal training data for estimating signal mean and background statistics to compute the Hotelling observer (HO) in digital breast tomosynthesis (DBT).

Methods:

Best among all linear observers, Hotelling observer (HO) is desired for many detection tasks. However, training data cost is usually high for obtaining prior information regarding the signal and data covariance. This limits the HO's utility in many practical situations. We propose that, in breast tomosynthesis images, it is feasible to estimate priors without using additional training data. The approach consists of obtaining the signal mean from a subtraction of noiseless reconstructions of a uniform background with and without signal and the noise power spectrum (NPS) computed from each breast tomosynthesis image itself, and implementing the HO in the Fourier domain. We call this approach no-additional-training Fourier HO (FHO). We compare this approach to the usual-training approach with pre-acquired training data to examine whether the resulting performances are comparable. We utilize a baseline model observer (MO) without prewhitening and an efficient MO with Laguerre-Gauss channelized HO (LG-CHO) to examine the efficiency of the no-additional-training FHO. Performance is evaluated in terms of lesion detectability using simulated DBT images corresponding to the UPENN breast phantoms with spherical lesions and two types of reconstruction.

Results:

The no-additional-training FHO yielded similar and even higher performance compared to the usual-training approach in all the simulated imaging conditions. The performance improvement over non-prewhitening MO was substantial and over LG-CHO was also moderate, indicating a higher efficiency of the no-additional-training FHO in detecting lesions from the simulated DBT images.

Conclusion:

The approach to estimate HO templates without additional training data shows excellent performance with simulated breast tomosynthesis data. Its efficiency will be further tested on virtual and real patient breast tomosynthesis image sets.

Ancillary