A framework for quantifying and leveraging uncertainty in pre-trained CT denoising model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Deep Learning for Medical Image Analysis

Background:

• Pre-trained deep learning models offer potential for low-dose computed tomography (CT) denoising.
• Accurate estimation and utilization of total uncertainty (aleatoric and epistemic) are crucial for reliable AI deployment in medical imaging.
• Existing models often lack robust uncertainty quantification, limiting their clinical utility and trustworthiness.

Purpose of the Study:

• To develop an architecture-agnostic framework for estimating, calibrating, and leveraging total uncertainty in pre-trained deep learning denoising models for low-dose CT.
• To improve the diagnostic image quality and lesion detectability in low-dose CT scans.
• To establish a foundation for performance monitoring, deployment optimization, and trustworthiness of AI models in medical imaging.

Main Methods:

• Developed a framework estimating aleatoric uncertainty via physics-based inference-time augmentation and epistemic uncertainty using training-free Monte Carlo dropout.
• Implemented non-parametric re-calibration to enhance uncertainty calibration, followed by adaptive local fusion (ALF) guided by the local mean-to-uncertainty ratio.
• Validated the framework on U-net and ResNet models across diverse CT datasets, assessing uncertainty with NRMSE/NCE and image quality/lesion detectability using SSIM and a deep learning model observer.

Main Results:

• The framework achieved accurate uncertainty quantification and calibration, with NRMSE in [1.2%, 2.4%] and NCE in [0.9%, 2.2%].
• ALF demonstrated comparable or reduced noise levels compared to original models, with significant improvements in lesion structural fidelity (SSIM) and detectability (p<0.05).
• Specific improvements included up to 69.7% noise reduction for lung nodules and up to 13.2% enhanced detectability for liver metastases.

Conclusions:

• The developed framework effectively quantifies and utilizes total uncertainty to enhance diagnostic image quality in low-dose CT using pre-trained denoising models.
• This approach facilitates improved performance monitoring, optimized deployment, and increased trustworthiness of AI in medical imaging applications.
• The architecture-agnostic nature of the framework allows broad applicability across various deep learning denoising models for CT.