M Lago*, A Badano, Food & Drug Administration, Silver Spring, MD

• 
  M Lago

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TU-J430-BReP-F1-4 (Tuesday, 7/12/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 1

Purpose: In medical imaging, there is a need to expand model observers to more realistic scenarios such as search tasks. As opposed to simple controlled synthetic noise backgrounds and location-known signals, training for search models in realistic images (e.g., breast imaging) introduces new challenges: backgrounds have non-stationary statistical properties and can vary significantly. In this research, we study how four model observers for search perform in backgrounds with varying noise correlation.

Methods: We generated 1000 white noise images that were convolved with a Gaussian filter of varying standard deviation. The filter's standard deviation increased from top to bottom. We trained each model observer (Non-Prewhitening, NPW; Channelized Hotelling Observer, CHO; Filtered Channel Observer, FCO; and Prewhitening Matched Filter, PWMF) only on only one of these correlations. On the testing phase, we convolved their templates with the whole image. Finally, we calculated the response for each trial by taking the maximum template response on each pixel when searching for a specific signal.

Results: When training is performed with a specific noise correlation but tested on different noise statistics, model responses on search tasks are unstable on areas in which they have not been trained. This is prevalent in images with high background statistical differences (e.g., breast imaging). One solution could be to have different models trained on different areas of the breast (depending on the glandularity) or significantly increase the training set.

Conclusion: We show that varying background statistics can lead to unexpected results when using a scanning model observer for search. We can see a similar behavior on more realistic and inhomogeneous backgrounds such as breast imaging. The two most tuned models (FCO and PWMF) suffer from an over-fitted training template. For search tasks in high-variable backgrounds, less tuned model observers seem to better tolerate statistical variations in the background.

Keywords

Observer Performance, Breast

Taxonomy

Not Applicable / None Entered.

Contact Email