A Naive-Bayes model observer for a human observer in detection, localization and assessment of perfusion defects in SPECT

In medical imaging, it is widely accepted that image quality should be assessed through the performance of human observers at some diagnostic task. For these evaluations, mathematical algorithms known as model observers (MOs) are often used as a substitute for human observers in early stages of image reconstruction algorithm development. For SPECT-MPI (myocardial perfusion imaging), diagnostic tasks involve detection, localization and assessment of regions with abnormal myocardial perfusion. In this work we propose a new MO for these tasks. The proposed MO is based on a machine-learning algorithm known as Naive-Bayes classifier (NB-MO). In the proposed approach, NB-MO is applied over a set of image features extracted from SPECT polar maps, aiming to predict perfusion scores given by humans for each myocardium region (segment). Next we compute average MO performance by first directly predicting individual human observer scores, and then using multi reader alternative free-response analysis (AFROC) in the same fashion as used on human observers. Our human observer study was comprised of five experienced (physicians) readers who scored location and severity of perfusion defects in 179 simulated SPECT-MPI cases and two different reconstruction methods (FBP and OSEM). The presented results show good performance agreement between humans and the proposed NB-MO, as well as excellent generalization properties of NB-MO between different reconstruction methods.