Channelized hotelling observers for detection tasks in multi-slice images

We investigate numerical observers for signal detection in volumetric imaging data sets viewed in stack browsing mode. Three types of multi-slice CHO (msCHO) as well as a single-slice CHO (ssCHO) are considered. We study the influence of signal size on model observer performance for detection of exactly known signals in three-dimensional (3D) non-Gaussian lumpy backgrounds (LB). The size of the signal is varied separately in the "frontal" or "coronal" plane (xy-plane) and in the "sagittal" plane (z-direction). METHODS The three msCHO designs: type a (msCHOa), type b (msCHOb) and type c (msCHOc), differ in how they treat the channelized slice data to infer the classification decision for the image. In one case, applied to msCHOa and msCHOb, the observers first build a test statistic for each slice and then use these to estimate the final test statistic for the image. The first step is modeled by the regular 2D-CHO applied on each slice in the stack and by the 1D-HO applied on the array of slice test statistics in the following step. For msCHOa, a separate 2D-CHO template is built for each position of the slice, while the msCHOb applies the same templates on multiple adjacent slices. In the other case, applied to msCHOc, the observers build their final statistics for the multi-slice image directly using the channelized slice data with no intermediate "scoring" on the slice level. Hence, the model applies the 1D-HO directly on the concatenated channelized slice data. Finally, the ssCHO corresponds to the conventional 2D-CHO applied on the central slice of the signal. CONCLUSION In general, the results of our study suggest higher observer performance of msCHO design compared to the ssCHO which confirms the expected benefit from using the data of multiple image slices in contrast to the single slice only. Due to its least restrictive assumptions among three types of msCHO models, type a may be applied in the greatest range of detection tasks. When applicable, type b achieves the highest performance, especially if the signal is spread across fewer slices. Not surprisingly, the results indicate significant dependency of the detection performance with regard to the signal characteristics, especially the signal size in xy-plane and signal spread across z-direction, but also with regard to the relation between the signal and the background structure. It is part of our future research plans to pursue a study with psychovisual experiments to compare the models and human observer performance.