A Feasibility Study of Domain-aware Deep Learning Ideal Model Observer

Assessment of medical image quality is a task historically performed by human experts, yet due to high cost and time constrains, algorithms called Model Observers (MO) have been developed to perform this task instead. Convolutional Neural Network (CNN) based MOs have previously been implemented for defect localization in medical images. Although effective, this approach lacks domain awareness - the ability to discern the unseen data types (i.e. different statistical characteristics) compared to the data it was trained on. In this paper, we are proposing an approach for defect localization which aims to achieve the accuracy of the ideal MO, as well as reconstruct the input image, for the purpose of achieving domain-awareness. By developing a domain-aware algorithm, we model a clinician's behavior, who can decide whether an image is familiar, and judge their own reliability in this context. Preliminary results show the algorithm's ability to successfully perform defect localization task alongside reconstructing the original input image. Additionally, we demonstrate a strong correlation between the defect localization accuracy and its mean squared error (MSE). These results reveal that a CNN based MO can be trained to perform both defect localization and image reconstruction simultaneously by minimizing two MSE functions - one for each task, instead of having to train two models which perform these tasks separately.