Related Experiment Video
Updated: Jun 21, 2025
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
42.6K
Unsupervised Domain Adaptation for Low-Dose CT Reconstruction via Bayesian Uncertainty Alignment.
Summary
This study introduces a novel probabilistic framework for low-dose computed tomography (LDCT) reconstruction, improving image quality and reliability in clinical settings by addressing domain variations and enhancing uncertainty quantification.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computational Imaging
Background:
- Low-dose computed tomography (LDCT) aims to reduce radiation exposure while maintaining image quality.
- Deep learning (DL) models for LDCT reconstruction struggle with performance degradation due to domain shifts between training and testing data.
- Unsupervised domain adaptation (UDA) methods align data distributions but often neglect uncertainty quantification and can cause content mismatch.
Purpose of the Study:
- To develop an advanced UDA method for LDCT reconstruction that incorporates uncertainty quantification.
- To address content mismatch issues in existing UDA techniques for cross-patient LDCT reconstruction.
- To improve the reliability and performance of DL-based LDCT reconstruction in diverse clinical scenarios.
Main Methods:
- A probabilistic reconstruction framework is proposed for joint discrepancy minimization in latent and image spaces.
- Bayesian uncertainty alignment is introduced in the latent space to reduce the epistemic gap and uncertainty in target domain data.
- Sharpness-aware distribution alignment (SDA) is employed in the image space to match second-order statistics and ensure image sharpness.
Main Results:
- The proposed method demonstrates superior quantitative and visual performance compared to existing methods on simulated and clinical LDCT datasets.
- Bayesian uncertainty alignment effectively reduces uncertainty in target domain data, leading to better reconstruction.
- Sharpness-aware distribution alignment ensures reconstructed target domain images possess comparable sharpness to normal-dose CT (NDCT) images.
Conclusions:
- The novel probabilistic UDA framework significantly enhances LDCT image reconstruction quality and reliability.
- Integrating uncertainty quantification and sharpness-aware alignment overcomes limitations of previous UDA methods in medical imaging.
- This approach holds promise for robust and dependable AI-driven medical image analysis in clinical practice.
