Artificial Intelligence Based Parametric X-Ray Imaging

Abstract

Spectral Photon-Counting Computer Tomography (SPCCT) offers the capability of exploring the difference in material attenuation of X-rays under different energy bins. This provides valuable spectral information that can aid in accurate material identification. This research project aims at developing a deep-learning approach for material identification in SPCCT to distinguish and quantify constituent materials within imaged object into muscle, adipose, different concentrations of iodine and gold, and different densities of hydroxyapatite. The study begins with the creation of a comprehensive SPCCT materials dataset using the QRM phantom comprising 61 meticulously annotated images per energy bin. The dataset serves as a foundational resource for AI model training and validation. Building on this dataset, the nnU-Net, a state-of-the-art DL framework for medical image segmentation, is employed to train and test a model for material identification. The model demonstrated robust performance, particularly for high-density materials, achieving an average Dice coefficient of 79.9% across all classes. Despite some challenges in distinguishing materials with similar pixel intensities, the model’s promising performance highlights its potential for clinical applications. The final phase of the research involved comparing the nnU-Net DL model with the MARS-MD algorithm, a conventional material identification method. The nnU-Net model outperformed the MARS-MD algorithm, particularly in real-world preclinical scenarios, demonstrating its generalizability and robustness. The potential impact of this research extends to multiple domains, including medical diagnostics, industrial quality control, and materials science research. This research advances spectral CT imaging capabilities, leading to enhanced tissue characterization, material classification, and non-destructive material property characterisation.

Date of Award	1 Jul 2024
Original language	American English
Supervisor	Maalej (Supervisor)