Tunable neural networks for CT image formation.

Area of Science:

• Medical imaging
• Computational imaging
• Deep learning

Background:

• CT image quality optimization balances variance and bias, controlled by filter cutoff or regularization strength.
• Deep neural networks (DNNs) lack tunable control over output image properties, typically minimizing mean squared error without trade-off management.
• Existing methods do not offer granular control over the variance-bias trade-off in DNN-based image reconstruction.

Purpose of the Study:

• To propose a novel loss function, weighted covariance and bias (WCB), for controlling DNN output image properties.
• To enable tunable control over the trade-off between variance and bias in neural network image reconstruction.
• To demonstrate task-specific optimization of image quality for DNNs.

Main Methods:

• Developed the weighted covariance and bias (WCB) loss function using multiple noise realizations for separate variance and bias penalty weighting.
• Employed spatial frequency domain penalties for targeted image feature penalization.
• Evaluated the method using simulations with digital anthropomorphic phantoms and CT measurement physics.

Main Results:

• The WCB loss function provides superior control over variance-bias trade-offs compared to mean squared error.
• WCB allows targeted control of specific image properties: variance, bias, spatial resolution, and noise correlation.
• Optimized WCB weights for a spiculated lung nodule shape discrimination task, demonstrating task-specific performance enhancement.

Conclusions:

• The WCB loss function offers a new paradigm for controlling DNN image properties in medical imaging.
• This method enables precise tuning of image quality attributes for improved diagnostic performance.
• WCB facilitates optimization of DNN outputs for specific clinical tasks, advancing medical image analysis.