4D Mathematical Observer Models for the Task-Based Evaluation of Gated Myocardial Perfusion SPECT Images

摘要

We developed and evaluated two 4D mathematical observer models for the detection of motion defects in simulated 4D gated myocardial perfusion (GMP) SPECT images, and compare predictions of the models with results from human observers. For the first model, we extended the application of conventional channelized Hotelling observer (CHO) method which uses 2D spatial channels to 4D motion images. For the second model, we developed a set of 'space-time channels' which includes a set of 2D spatial channels and an additional 1D channel in the time direction. To evaluate the models, we used the 4D eXtended CArdiac Torsro (XCAT) phantom and created 16 gates/cardiac cycle with normal myocardial perfusion but a regional wall motion defect (RWMD). Simulated GMP SPECT projection data were generated using an analytical projector and reconstructed using FBP followed by spatial and temporal low-pass filtering with various cut-off frequencies. For the first model, we extracted 2D slices from each time frame containing the defect center and produced feature vectors using the 2D channels. Then, they were stacked for all the time frames to which the ROC methodology was applied. For the second model, the extracted 2D image slices were reorganized into a set of cine 2D images and the space-time channels were applied to produce space-time feature vectors to which the ROC was applied. For both models, the rating data were analyzed and areas under the ROC curve (AUCs) were computed. The resulting AUC values of the second model showed better agreement to those previously obtained from a human observer study viewing the same data than those from the first model, and positive correlation in ranking between them was found. We have demonstrated that RWMD detection was affected by post-filtering. The AUC values from the proposed 4D space-time CHO model show good agreement with that from a human observer study in detecting a RWMD in the same set of simulated 4D GMP SPECT images.