Enhancing signal detectability in learning-based CT reconstruction with a model observer inspired loss function.

Area of Science:

• Medical imaging
• Computational imaging
• Artificial intelligence in healthcare

Background:

• Current deep neural networks for sparse-view CT reconstruction often use pixel-wise losses (e.g., mean-squared error).
• These methods can obscure small, low-contrast features vital for accurate medical diagnosis and screening.
• There is a need for improved reconstruction techniques that preserve subtle image details.

Approach:

• A novel training loss function inspired by the model observer framework was developed.
• This loss function aims to enhance the detectability of weak signals within the reconstructed images.
• The approach was evaluated on synthetic sparse-view breast CT data.

Key Points:

• The proposed loss function significantly improves the detectability of low-contrast signals compared to traditional pixel-wise losses.
• Small, diagnostically relevant features are better preserved in the reconstructed images.
• The model observer-inspired loss offers a promising direction for enhancing sparse-view CT reconstruction quality.

Conclusions:

• The novel training loss effectively addresses the limitation of pixel-wise losses in deep learning-based CT reconstruction.
• This method holds potential for improving the diagnostic accuracy of sparse-view CT imaging, particularly in breast cancer screening.
• Further research can explore the application of this loss function to other medical imaging modalities and sparse-data scenarios.