Application of learned ideal observers for estimating task-based performance bounds for computed imaging systems

The performance of the ideal observer (IO) acting on imaging measurements has long been advocated as a figure-of-merit (FOM) to guide the optimization of imaging systems. For computed imaging systems, the performance of the IO acting on imaging measurements also sets an upper bound on task-performance that no image reconstruction method can transcend. As such, estimation of IO performance can provide valuable guidance when designing data-acquisition techniques by enabling the identification of designs that will not permit the reconstruction of diagnostically useful images for a specified task - no matter how advanced the reconstruction method is or plausible the reconstructed images appear. While such data space IO analyses are known conceptually, they have generally remained infeasible to widely implement. In this work, convolutional neural network (CNN) approximated IOs (CNN-IOs) are investigated for estimating the performance of data space IOs for the purpose of guiding hardware and data-acquisition designs and establishing task-based performance bounds for image reconstruction.

Numerical studies that utilized a stylized breast X-ray computed tomography test bed are conducted to validate and demonstrate the approach. Signal-known-statistically and background-known-statistically (SKS/BKS) binary signal detection and discrimination tasks are addressed and the impact of the number of views and beam intensities on IO performance is investigated as a case study. The image space CNN-IO performance is also computed by use of images reconstructed by both U-Net and FBP reconstruction methods and compared to the corresponding data space CNN-IO performance to assess task-related information loss.

For all considered cases, task-performance bounds were established by use of the data space CNN-IO performance. A comparison of the data space and image space CNN-IO performances quantified the task-relevant information loss induced by the considered image reconstruction methods. Moreover, the U-Net reconstructed images possessed improved traditional metrics compared to those produced by the FBP method but resulted in lower image space CNN-IO performance. This demonstrates that traditional IQ measures can be misleading if task-performance is of ultimate interest.

This work confirms that recent developments in learning-based IO approximation methods can enable the ranking of data-acquisition designs based on optimal task-performance with consideration of object variability. The work also demonstrates that such methods can enable estimation of task-based performance bounds for image reconstruction.