Probabilistic U-Net model observer for the DDC method in CT scan protocol optimization.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Optimizing Computed Tomography (CT) imaging involves balancing radiation dose and image quality, a complex task exacerbated by technological advancements.
• Current methods for evaluating image quality, such as Difference-Detailed Curves (DDC), rely on human observer studies that are time-consuming and suffer from variability.
• Machine learning (ML) offers potential solutions to automate and standardize image quality assessment in CT.

Purpose of the Study:

• To develop and validate a machine learning-based model observer for optimizing CT imaging protocols.
• To overcome the limitations of human observer studies, including labor intensity and inter/intra-observer variability.
• To predict Difference-Detailed Curve (DDC) distributions for improved CT protocol optimization.

Main Methods:

• A U-Net architecture and Bayesian methodology were employed to create a ML-based model observer.
• Gaussian Process-based noise modeling was used for image preprocessing to ensure robustness against object spatial arrangement.
• Gradient-weighted class activation mapping (Grad-CAM) was utilized for model interpretability.
• Beta regression principles informed the Bayesian methodology to derive a performance metric ('effective number of observers').

Main Results:

• The proposed ML model observer achieved well-calibrated probabilistic predictions by training on diverse observer data, quantifying observer variability.
• The Bayesian methodology provided a performance metric, quantifying the model observer's strength as an 'effective number of observers'.
• The framework successfully predicted DDC distributions by applying thresholds to inferred probabilities, enabling efficient CT protocol optimization.

Conclusions:

• The developed ML model observer provides a robust and efficient alternative to traditional human observer studies for CT image quality assessment.
• This approach effectively quantifies observer variability and aids in optimizing CT protocols for both dose and image quality.
• The framework offers a scalable solution for enhancing the accuracy and reliability of CT imaging procedure optimization.