Metal Artifacts Reduction in CT Scans using Convolutional Neural Network with Ground Truth Elimination.

Metal artifacts are very common in CT scans since metal insertion or replacement is performed for enhancing certain functionality or mechanism of patient's body. These streak artifacts could degrade CT image quality severely, and consequently, they could influence clinician's diagnosis. Many existing supervised learning methods approaching this problem assume the availability of clean images data, images free of metal artifacts, at the part with metal implant. However, in clinical practices, those clean images do not usually exist. Therefore, there is no support for the existing supervised learning based methods to work clinically. We focus on reducing the streak artifacts on the hip scans and propose a convolutional neural network based method to eliminate the need of the clean images at the implant part during model training. The idea is to use the scans of the parts near the hip for model training. Our method is able to suppress the artifacts in corrupted images, highly improve the image quality, and preserve the details of surrounding tissues, without using any clean hip scans. We apply our method on clinical CT hip scans from multiple patients and obtain artifact-free images with high image quality.