Abstract:
This thesis presents a novel approach to enhancing material decomposition (MD) in
medical imaging using the MARS CT scanner. Traditional MD techniques are plagued
by issues like computational complexity, frequent re-calibration needs, and cross-talk
between materials with similar densities, which complicate accurate classification and
quantification. This study effectively addresses these challenges by developing and
implementing the advanced ResUNet++ model. ResUNet++ leverages nested convolutional blocks, dense skip connections, and residual layers to maintain high image resolution and improve feature propagation and reuse, mitigating the vanishing gradient
problem. Comprehensive testing on both standard and blind datasets demonstrated the
model’s robustness, significantly reducing noise and cross-talk and leading to marked
improvements in classification accuracy and quantification precision across elements
like Gold (Au), Calcium (Ca), Gadolinium (Gd), Iodine (I), Lipid, and Water. Comparative analysis with the conventional MD method used in the MARS scanner highlights
the superior performance of ResUNet++, eliminating the need for pre-calibration. The
study concludes that ResUNet++ sets a new benchmark for MD accuracy, offering a robust solution for more reliable and efficient diagnostic procedures in medical imaging.
Future work will focus on testing the model on biological data and training it to adapt to
various protocols, ensuring broader applicability and enhanced performance in diverse
medical imaging scenarios.
