Using the Hotelling observer on multi-slice and multi-view simulated SPECT myocardial images

A simulation study was done to investigate a novel use of the Hotelling Observer. A two-layer model which included a Channelized Hotelling; Observer (CHO) followed by a Hotelling Observer (HO) was built to operate on three dimensional images containing multiple slices and multiple views similar to SPECT myocardial perfusion images. The simulated images were hearts with and without cold lesions, with noise and blur. For one lesion and each of 20 lesion locations, 1000 lesion-present and 1000 lesion-absent images were generated. These 40,000 images consisted a data set for that one lesion. Each image was reformatted into three rows, providing 10 adjacent short axis slices, 10 adjacent vertical long axis slices, and 10 adjacent horizontal long axis slices. The CHO was computed for each slice, giving 40,000 3×10 decision variable arrays. A HO was then applied on the decision variable arrays to obtain a multi-slice CHO-HO lesion detectability index. Three additional ensembles were generated for single slices through the lesion center, one short axis only, one vertical long axis only, and one horizontal long axis only. Applying the CHO to these ensembles gave lesion detectabilities for the short axis view, vertical long axis view, and horizontal long axis view respectively. An 10-fold cross validation was used to compute uncertainties in all detectabilities. Three data groups, total of 23 such sets with different lesion contrast, size, or shape were generated to compare the correlations between the performance of the multi-slice CHO-HO, the single-slice CHO, and human observers. The result suggests the ability of the multi-slice method to separate better these images into their correct classes as compared to the single-slice method, by combining information from multiple slices. The multi-slice method provides a single detectability scalar for studies in which the lesion could be present in multiple slices. It may be that a multi-slice CHO-HO model could more closely predict the performance of radiologists in tasks where the radiologist evaluates multi-slice and multi-view images