A hybrid training database and evaluation benchmark for assessing metal artifact reduction methods for X-ray CT...

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Metal artifacts in computed tomography (CT) significantly degrade image quality, hindering accurate diagnosis.
• Existing metal artifact reduction (MAR) algorithms lack a comprehensive, clinically relevant evaluation benchmark.
• The absence of artifact-free ground truth data and paired training datasets limits the development of advanced MAR techniques, especially deep learning-based methods.

Purpose of the Study:

• To propose a simulation framework for generating a large training database for deep learning-based MAR algorithms.
• To define a comprehensive evaluation benchmark for MAR algorithms using realistic simulated metal artifacts.
• To validate a framework for the realistic simulation of metal artifacts on clinical CT data.

Main Methods:

• Modeled clinical and generic CT scanner geometries using the CatSim CT simulator within the XCIST toolkit.
• Simulated 2D datasets to generate metal artifact scenarios for training deep learning algorithms, utilizing public CT databases.
• Validated the metal artifact simulation capability experimentally and defined a benchmark for clinically realistic scenarios, applied to numerical and deep learning MAR algorithms.

Main Results:

• The simulation tool demonstrated high accuracy, with mean CT number deviation below 2% compared to real data.
• Generated 14,000 simulated metal artifact scenarios across head, thorax, and pelvis regions.
• Defined a benchmark with metrics for CT number accuracy, noise, sharpness, and structural integrity, covering diverse clinical metal implant scenarios. Simulation tools and benchmark datasets were made publicly available.

Conclusions:

• Developed and distributed novel tools and datasets essential for the advancement of MAR algorithms.
• Established the first comprehensive evaluation benchmark for MAR, encompassing a wide array of clinically realistic metal artifact scenarios.
• Facilitated the development and rigorous evaluation of MAR algorithms, particularly deep learning approaches, through publicly available resources.