2D Lung Thickness Estimation from Chest X-Rays Using U-Net Regression Trained with Digitally Reconstructed Radiographs
Date of Award
5-5-2024
Degree Name
M.S. in Computer Engineering
Department
Department of Electrical and Computer Engineering
Advisor/Chair
Russell Hardie
Abstract
Chest X-rays (CXRs) are one of the most common medical imaging procedures providing two-dimensional (2D) images of three-dimensional (3D) density data regarding a patient’s chest. Computed tomography (CT) scans give a more extensive look at a desired location by utilizing X-rays to provide a 3D view of a specified area of the human body in slices. CT scans are quintessential for getting lung measurements as well as identifying and tracking lung cancer nodule growth within said lungs. Many different computer aided detection (CAD) systems have the ability to read CT scan data and assist medical professionals in outlining essential information within the lungs such as providing lung outlines, detecting lung nodules, and more. The ability for a CAD system to take advantage of lung thickness information would assist with bounding CAD systems but also providing algorithms with information that can assist in determining the likelihood of a nodule present in certain lung areas. In this approach, a method by which CT scans are converted into synthetic CXRs is introduced. In the process of generating these synthetic CXRs, a corresponding set of relative 2D lung thickness values is generated for each pixel in which the lung exists within a scan as a beam travels through the lung from front to back. A regression neural network (RNN) is then created based on U-Net architecture to train a model to predict the relative thickness of the lungs using the data from the synthetic CXR generation. CT scans from the Lung Image Database Consortium-Image Database Research Initiative (LIDC-IDRI) are used to generate synthetic CXRs and the associated lung thickness data. After the data has been processed, scaled, and augmented, it is used to train and test the U-Net RNN, which can predict relative lung thickness in other synthetic CXRs with an overall mean absolute error (MAE) of 0.0301 and an overall mean squared error (MSE) of 0.0047.
Keywords
DRR, U-Net Regression, Lung Thickness, Computer Aided Detection Algorithms, Deep Learning
Rights Statement
Copyright 2024, author
Recommended Citation
Marsh, John J., "2D Lung Thickness Estimation from Chest X-Rays Using U-Net Regression Trained with Digitally Reconstructed Radiographs" (2024). Graduate Theses and Dissertations. 7594.
https://ecommons.udayton.edu/graduate_theses/7594
