Personalized numerical observer

It is widely accepted that medical image quality should be assessed using task-based criteria, such as humanobserver (HO) performance in a lesion-detection (scoring) task. HO studies are time consuming and cost prohibitive to be used for image quality assessment during development of either reconstruction methods or imaging systems. Therefore, a numerical observer (NO), a HO surrogate, is highly desirable. In the past, we have proposed and successfully tested a NO based on a supervised-learning approach (namely a support vector machine) for cardiac gated SPECT image quality assessment. In the supervised-learning approach, the goal is to identify the relationship between measured image features and HO myocardium defect likelihood scores. Thus far we have treated multiple HO readers by simply averaging or pooling their respective scores. Due to observer variability, this may be suboptimal and less accurate. Therefore, in this work, we are setting our goal to predict individual observer scores independently in the hope to better capture some relevant lesion-detection mechanism of the human observers. This is even more important as there are many ways to get equivalent observer performance (measured by area under receiver operating curve), and simply predicting some joint (average or pooled) score alone is not likely to succeed.