Approximating the Hotelling observer with autoencoder-learned efficient channels for binary signal detection tasks

PurposeThe objective assessment of image quality (IQ) has been advocated for the analysis and optimization medical imaging systems. One method computing such IQ metrics is through a numerical observer. The Hotelling observer (HO) optimal linear observer, but conventional methods obtaining HO can become intractable due to large sizes or insufficient data. Channelized are sometimes employed in circumstances approximate HO. performance channelized varies, with different superior others depending on conditions detection task. A using an AE presented implemented across several tasks characterize its performance.ApproachThe process training demonstrated be equivalent developing set channels approximating efficiency learned AE-channels increased by modifying loss function incorporate task-relevant information. Multiple binary involving lumpy breast phantom backgrounds varying dataset considered evaluate proposed compare current state-of-the-art methods. Additionally, ability generalize images outside investigated.ResultsAE-learned have comparable other channel studies generalization studies. Incorporating cleaner estimate signal task also significantly improve method, particularly datasets fewer images.ConclusionsAEs capable learning efficient resulting significant increase small when incorporating prior holds promising implications future assessments technologies.