Development of a denoising convolutional neural network-based algorithm for metal artifact reduction in digital...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Image Reconstruction Techniques

Background:

• Metal artifacts significantly degrade image quality in digital tomosynthesis (DT) for arthroplasty.
• Existing metal artifact reduction (MAR) methods often struggle with complex metal objects and specific imaging modalities.

Purpose of the Study:

• To develop and evaluate a novel denoising convolutional neural network metal artifact reduction hybrid reconstruction (DnCNN-MARHR) algorithm.
• To assess the efficacy of DnCNN-MARHR in reducing metal artifacts in DT projection data for arthroplasty.

Main Methods:

• Implementation of a DnCNN-MARHR algorithm using a training network to estimate residual images from projection data.
• Hybrid reconstruction combining back projection and maximum likelihood expectation maximization (MLEM).
• Comparative analysis against dual-energy material decomposition reconstruction algorithm (DEMDRA), MLEM, filtered back projection (FBP), and SART-TV with MAR.

Main Results:

• DnCNN-MARHR demonstrated adequate performance in artifact spread function (ASF) analysis, comparable to DEMDRA.
• The algorithm achieved superior MAR compared to conventional methods, with improved image homogeneity shown in texture analysis.
• Mean square error analysis indicated DnCNN-MARHR yielded the smallest virtual monochromatic (VM) difference.

Conclusions:

• The proposed DnCNN-MARHR algorithm offers a robust solution for metal artifact reduction in DT imaging of arthroplasty.
• It effectively reduces artifacts, particularly in the longitudinal direction, without being influenced by metal type or causing tissue misclassification.