Mirmojarabian, Mohammadmahdi (2024) Overcoming limited field of view challenges in whole-body CT scans: a GAN approach for imputing truncated tissues. Masters thesis, Memorial University of Newfoundland.
Full text not available from this repository.
Abstract
One of the significant challenges in medical imaging is the limited field of view (FOV) in computed tomography (CT) scans, which often results in truncated tissues. This limitation can hinder accurate diagnosis and affect subsequent medical decisions, as it can lead to incorrect body composition (BC) measurements. While existing AI methods for FOV extension have shown promise in recovering truncated parts, they are limited to two-dimensional (2D) CT images. This thesis presents a two-stage method for extending the FOV in whole-body three-dimensional (3D) CT scans to address tissue truncation caused by limited FOV. In the first stage, a bounding box predicts the full body extent, allowing the FOV to be symmetrically extended. The second stage uses an image completion model (Recurrent Feature Reasoning Network (RFR-Net)) to fill in the missing tissue in truncated regions. To effectively train the model, we generated synthetic FOV truncation patterns from complete CT slices. The method was rigorously evaluated by comparing BC metrics, such as skeletal muscle and fat, before and after truncation correction. Our results demonstrate that the model successfully restores missing anatomical structures and significantly reduces errors in BC measurements. Specifically, the Mean Absolute Percentage Error (MAPE) for skeletal muscle volume was reduced from 39.28% to 6.56%, for fat volume from 44.55% to 7.35%, and for bone volume from 32.23% to 7.68%. Furthermore, the pixel-wise hounsfield unit (HU) performance metrics showed promising results, with an overall Structural Similarity Index (SSIM) of 0.9871. This approach offers a practical solution for enhancing CT scan quality in cases of limited FOV, thereby improving clinical assessments and diagnostics in medical imaging applications.
| Item Type: | Thesis (Masters) |
|---|---|
| URI: | http://research.library.mun.ca/id/eprint/16802 |
| Item ID: | 16802 |
| Additional Information: | Includes bibliographical references (pages 56-67) -- Restricted until November 22, 2025 |
| Keywords: | generative adversarial networks (GANs), computed tomography (CT) imaging, medical image extension, field of view (FOV) truncation, artificial intelligence in healthcare |
| Department(s): | Science, Faculty of > Computer Science |
| Date: | September 2024 |
| Date Type: | Submission |
| Digital Object Identifier (DOI): | https://doi.org/10.48336/1rkt-j578 |
| Library of Congress Subject Heading: | Diagnostic imaging; Tomography; Artificial intelligence--Medical applications; Neural networks (Computer science); Machine learning |
Actions (login required)
 |
View Item |
