Trainable joint bilateral filters for enhanced prediction stability in low-dose CT

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Low-dose computed tomography (CT) requires effective denoising to maintain image quality while reducing patient radiation exposure.
• Deep learning (DL) methods show promise but struggle with generalization beyond training data in clinical settings.
• Conventional denoising algorithms lack the model capacity of DL approaches.

Purpose of the Study:

• To develop a hybrid denoising approach that combines the strengths of DL and conventional methods for robust low-dose CT image enhancement.
• To improve the generalization capabilities of DL-based denoising algorithms in medical imaging applications.
• To validate the proposed method on diverse CT datasets, including those with metal artifacts and different anatomical regions.

Main Methods:

• A hybrid denoising pipeline integrating trainable joint bilateral filters (JBFs) with a convolutional DL network.
• The DL network is utilized for feature extraction and predicting a guidance image for the JBFs.
• The approach was trained on abdomen CT scans without metal implants and tested on scans with metal implants and head CT data.

Main Results:

• The hybrid method significantly improved denoising performance compared to standalone DL denoisers (RED-CNN/QAE) on challenging datasets.
• Performance gains included reductions in RMSE and increases in PSNR, particularly in metal artifact regions (up to 82%) and on head CT data (up to 78%).
• The trainable JBFs effectively constrained the error bounds of the deep neural networks.

Conclusions:

• The proposed hybrid denoising approach enhances the generalization and reliability of DL-based methods for low-dose CT.
• This technique facilitates the clinical adoption of DL denoisers by improving robustness across different imaging scenarios.
• The integration of trainable JBFs offers a promising direction for developing more dependable medical image denoising solutions.