TU‐EF‐204‐11: Impact of Using Multi‐Slice Training Sets On the Performance of a Channelized Hotelling Observer in a Low‐Contrast Detection Task in CT

Abstract

Purpose:To investigate using multiple CT image slices from a single acquisition as independent training images for a channelized Hotelling observer (CHO) model to reduce the number of repeated scans for CHO‐based CT image quality assessment.Methods:We applied a previously validated CHO model to detect low contrast disk objects formed from cross‐sectional images of three epoxy‐resin‐based rods (diameters: 3, 5, and 9 mm; length: ∼5cm). The rods were submerged in a 35x 25 cm2 iodine‐doped water filled phantom, yielding‐15 HU object contrast. The phantom was scanned 100 times with and without the rods present. Scan and reconstruction parameters include: 5 mm slice thickness at 0.5 mm intervals, 120 kV, 480 Quality Reference mAs, and a 128‐slice scanner. The CHO's detectability index was evaluated as a function of factors related to incorporating multi‐slice image data: object misalignment along the z‐axis, inter‐slice pixel correlation, and number of unique slice locations. In each case, the CHO training set was fixed to 100 images.Results:Artificially shifting the object's center position by as much as 3 pixels in any direction relative to the Gabor channel filters had insignificant impact on object detectability. An inter‐slice pixel correlation of >∼0.2 yielded positive bias in the model's performance. Incorporating multi‐slice image data yielded slight negative bias in detectability with increasing number of slices, likely due to physical variations in the objects. However, inclusion of image data from up to 5 slice locations yielded detectability indices within measurement error of the single slice value.Conclusion:For the investigated model and task, incorporating image data from 5 different slice locations of at least 5 mm intervals into the CHO model yielded detectability indices within measurement error of the single slice value. Consequently, this methodology would Result in a 5‐fold reduction in number of image acquisitions.This project was supported by National Institutes of Health grants R01 EB017095 and U01 EB017185 from the National Institute of Biomedical Imaging and Bioengineering.