Ideal observers from a visual-search paradigm: localization tasks and performance figures of merit

Objectives: The channelized Hotelling (CH) observer is an ideal linear observer for fixed-location lesion detection. When applied in scanning mode for search tasks, the CH observer is tractable but also suboptimal and computationally expensive. We report on an alternative search-capable observer that is both optimal and efficient.

Methods: A free-response detection task employed a binary Hotelling discriminant constructed from extracted feature values at lesion-present and lesion-absent locations in a set of images. These locations resulted from a simulated visual-search (VS) process that identified local maxima in a specified set of feature maps. The search sensitivity can be adjusted with a lower threshold on the accepted maxima; herein, the threshold was 70% of the image maximum for each search feature. Our tested features were gradient and difference-of-Gaussian channels commonly used with the CH observer. Optimal classification performance was quantified by the Hotelling SNR [1] and then converted to an equivalent area under the ROC curve. Observer performance in the free-response study was compared with the performance in a conventional localization ROC (LROC) study where the task was to identify the maximally suspicious location in each image. These studies used sets of clinical Tc-99m SPECT lung images containing synthetic nodules. Four iterative reconstruction strategies were devised with different combinations of attenuation, scatter and resolution correction. For each strategy, RBI-EM [2] reconstructed volumes were generated for 15 pairings of iterations and Gaussian postfilter levels. The studies used 154 2D image slices (54 training and 108 test images) extracted from the volumes, with an equal number of lesion-absent and single-lesion-present cases. LROC data acquired from three lay human observers for a subset of the image sets were included in our analysis.

Results: The correlation coefficients between the model-observer LROC and free-response performances for the four reconstruction strategies ranged from 0.85 to 0.97. Agreement between the model and average human in the LROC study could be modified by adjusting the former’s search sensitivity.

Conclusion: Anthropomorphic model observers can be devised by applying thresholds within the VS framework of this ideal observer. The ideal model thus provides a theoretical basis for VS observers that have been recently proposed [3] as surrogates for human observers in clinically realistic diagnostic imaging tasks. Research Support: This work was supported in part by NIBIB grant R01-EB012070.