MRI Susceptibility Mapping of Brachytherapy Seeds with Deep Learning

Abstract

Contrast is an essential element of magnetic resonance imaging. Good contrast in an MR image is necessary to correctly differentiate between different tissue structures and make accurate diagnoses. However, objects with high magnetic susceptibility, like metallic objects, cause severe artifacts that interfere with operations and routine evaluations.

In this work, we present a deep learning-based method to undo these effects, with a focus on brachytherapy seeds. We train the network on synthetic data to generate positive contrast images from magnetic field maps for localizing the seeds from their surroundings. We evaluate the model on other synthetic data, then show that the proposed model exhibits generalization to real MRI data and outputs a result quickly. We compare its performance with another positive contrast algorithm for brachytherapy seeds to demonstrate the potential of the deep learning implementation.