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

Area of Science:

• Medical Imaging
• Computational Imaging
• Observer Performance Modeling

Background:

• The ideal observer (IO) is a benchmark for imaging system optimization, setting theoretical performance limits.
• Estimating IO performance guides data acquisition and identifies designs incapable of producing diagnostically useful images.
• Data space IO analysis is conceptually known but historically difficult to implement widely.

Purpose of the Study:

• To investigate convolutional neural network (CNN) approximated IOs (CNN-IOs) for estimating data space IO performance.
• To guide hardware and data acquisition design in computed imaging systems.
• To establish task-based performance bounds for image reconstruction methods.

Main Methods:

• Numerical studies using a stylized breast X-ray computed tomography test bed.
• Signal-known-statistically and background-known-statistically (SKS/BKS) binary detection and discrimination tasks.
• Comparison of data space CNN-IO performance with image space CNN-IO performance on images reconstructed by U-Net and Filtered Back Projection (FBP).

Main Results:

• Task-performance bounds were established using data space CNN-IO performance.
• Quantified task-relevant information loss from image reconstruction methods.
• U-Net images had better traditional metrics but lower image space CNN-IO performance than FBP, indicating traditional metrics can be misleading.

Conclusions:

• Learning-based IO approximation methods, like CNN-IOs, enable ranking of data acquisition designs based on optimal task-performance.
• These methods can estimate task-based performance bounds for image reconstruction.
• Traditional image quality measures may not reflect true task-based performance, highlighting the value of observer performance modeling.