Keywords: Computerized Phantoms, Image Registration, Convolutional Neural Networks
Abstract: Computerized phantoms play an important role in medical imaging research. They can serve as a gold standard for evaluating and optimizing medical imaging analysis, processing, and reconstruction methods. Existing computerized phantoms model anatomical variations through organ and phantom scaling, which does not fully capture the range of anatomical variations seen in humans. Here, we present a registration-based method for creating highly realistic and detailed anthropomorphic phantoms. The proposed registration method is built on the use of an unsupervised convolutional neural network (ConvNet) that warps the four-dimensional Xtended Cardiac-Torso (XCAT) phantom to a patient CT scan. The registration ConvNet iteratively optimizes an SSIM-based loss function for a given image pair without prior training. We experimentally show substantially improved image similarity of the generated phantom using the proposed method to a patient image.
Paper Type: both
Primary Subject Area: Application: Radiology
Secondary Subject Area: Image Registration
Paper Status: based on accepted/submitted journal paper
Source Code Url: https://github.com/junyuchen245/Fully_Unsupervised_CNN_Registration_Keras
Data Set Url: https://wiki.cancerimagingarchive.net/display/Public/NaF+Prostate https://olv.duke.edu/technologies/4d-extended-cardiac-torso-xcat-phantom-version-2-0/
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.