Development and clinical validation of domain-aware anthropomorphic model observers for task-based image quality assessment

In this project we aim to develop deep-learning anthropomorphic model observers (AMO) as a substitute for human observers (HO) in studies aimed to assess image quality, defined in a task- based fashion based on clinical diagnostic performance. An accurate AMO would allow fast and clinically relevant procedure for optimization of imaging system and algorithm designs. The proposed AMO methods aim to achieve good generalization, i.e., an AMO developed for images reconstructed by one algorithm (or scanner setup) should predict HO performance accurately for a different reconstruction algorithm (or different scanner setting). In other words, a desired AMO should be tolerant to domain shifts. It is further proposed to design AMO which is aware of its own limitations and type of images (or organs) it can be used on, so that it does not attempt to evaluate images it is unsuited to judge. To this end, we propose to augment the AMO with the capability of domain-awareness to recognize image domain shift, and to provide a mechanism that can be used to economize the use of HO studies by more efficiently adapting the AMO to newly available datasets This project will also aim to overcome common concerns about deep learning being a “black- box” by leveraging methods from visual psychophysics, such as reverse-correlation. The advantage of this approach is that the identical procedure had been applied to HO, thus allowing a direct comparison and interpretation of the resulting receptive fields. To accurately train, test and clinically validate AMO we will develop a rich, image data sets, annotated by experts and non-experts, mimicking CT imaging of the liver. Therefore, a key component of the project will be development of extensive annotated datasets ranging from numerical simulation, acquisition from 3D printed phantoms to clinical patient data. For image acquisition we will use simulated, virtual CT and clinical (physical) CT scanning followed by several reconstruction methods. We choose to use liver CT imaging as the development platform for AMOs in low-contrast lesion localization and discrimination tasks. The chosen application is significant by itself because, despite years of research, there is still no systematic framework for optimization of the many factors in imaging system design for clinical diagnostic tasks. We believe that the proposed domain-aware AMO approach and AMO interpretability methodology will have a substantial and lasting effect not only on the methodology of AMO development and use, but also on general deep-learning applications in medical imaging.