Constructing efficient channels for ideal observers using the conjugate gradient method

Task-based assessment of image quality is critically important for the design and optimization of medical imaging systems. Ideal observers, including the Bayesian ideal observer (IO) and the ideal linear observer, i.e., the Hotelling observer (HO), provide objective figures of merit that quantify system performance on signal detection tasks. However, the application of ideal observers to high-dimensional image data is often computationally intractable. Channel mechanisms provide an effective framework for dimensionality reduction that can facilitate the computation of ideal observers. We present a conjugate gradient (CG)-based method to construct efficient channels for approximating the IO and HO performance.

The channels are constructed using the CG method associated with the optimization of linear discriminant for signal detection tasks. The proposed method is capable of incorporating prior information about the objects being imaged. Numerical studies are conducted for a binary signal detection task involving type I lumpy backgrounds. The proposed CG-based channels are compared with partial least squares (PLS) channels in terms of their ability to preserve task-relevant information for approximating the IO and HO performance. The effects of channel number and training dataset size on signal detection performance are systematically investigated.

For the considered signal-known-exactly and background-known-statistically binary detection task, the CG-based channels provide efficient low-dimensional representations of images that preserve task-relevant information and enable accurate estimation of the IO and HO performance. Compared with the PLS channels, the proposed method that incorporates prior information about the objects being imaged produces cleaner channels and improved detection performance, particularly under limited training data conditions.

The proposed CG-based channelization approach provides a principled and efficient way for extracting task-relevant features that enable accurate approximation of the IO and HO performance. By incorporating prior information about the objects being imaged, the proposed method can effectively construct efficient channels using a limited training dataset.