Introduction of a Pixel-Specific Model Observer to Accommodate Spatially Complex Clinical Devices As Test Objects
Abstract
Purpose
Existing spatial domain implementations of Hotelling observers used to assess imaging system performance provide a single SNR figure of merit associated with a symmetric test object. The purpose of this work was to develop and validate a pixel-specific model observer which directly estimates the snr2 contribution of each pixel to test object SNR2, thus making the approach directly applicable to spatially complex clinical devices as test objects.
Methods
Test object SNR was estimated using three methods. One method directly used pixel values as input to an ‘ideal’ Hotelling observer model (HO). The second method was a Gabor function channelized Hotelling observer (CHO). Finally, a pixel-specific observer (PO) directly estimated the snr2 contribution of each pixel to test object SNR2. For each method, the multiplicative effect of finite sampling bias was corrected using established theory. Unprocessed sample images (n=1200) of 0.5 to 2 mm iodine disks were acquired using an X-ray angiography system and submitted to the three variants of multivariate statistical methods. The PO method was additionally applied to images of clinical devices acquired using 0.4 and 0.9 mm focal spots.
Results
Iodine disk SNR by the HO method was in the range 1.8 to 10.6. The ratio of CHO to HO estimates was in the range 0.95 to 1.06 and the ratio of PO to HO estimates was in the range 0.8 to 1.0. Both the CHO and PO methods provided SNR estimates using substantially fewer samples than required for the HO. Pixel-specific snr2 maps of clinical devices were sensitive to 0.4 vs. 0.9 mm focal spot blur.
Conclusion
The three methods tested resulted in consistent estimates of iodine disk SNR. The novelty of the PO method is that the pixel-specific snr2 estimates can be used to estimate SNR2 of an entire test object or any portion thereof.