Numerical observer for cardiac motion assessment using a linear discriminant

In this work, we aim to develop a numerical observer for cardiac motion assessment in cardiac-gated single photon emission computed tomography (SPECT) imaging. In clinical imaging, diagnostic performance of a human observer for a specific diagnostic task is the gold standard of image quality assessment. However human studies are impractical and costly therefore the use of a numerical observer, a mathematical model that emulates human observer, is nowadays a common practice. To this end, the channelized Hotelling observer is a numerical observer used to assess image quality in myocardial perfusion defect detection tasks. However, cardiac motion, another important diagnostic feature for cardiac imaging, has not been used to evaluate image quality. Thus, we propose a new numerical observer aiming to predicting human observers´ performance in detection of cardiac motion defects. The proposed observer uses regional features extracted by our previously introduced and evaluated motion estimation technique, which relies on a deformable mesh model. Next, these features are used in a linear discriminant model in order to predict human performance. Note that this work does not rely on a channelized model of the human visual system for feature extraction as the CHO does. Instead we use a deformable mesh model capturing the relevant regional motion information. For evaluation, we have used a simulated cardiac motion defect and performed comparison with a human observer study for two types of post-reconstruction filtering strategies. Observers´ comparison suggests that the proposed methods can accurately predict human detection performance in assessment of cardiac motion defect.