UCLA

Quantitative imaging in CT promises objective descriptors of disease from medical images, but clinical translation is limited by sensitivity to CT acquisition and reconstruction protocol. The same object imaged under different scanner models, reconstruction kernels, or dose levels can yield substantially different measurements, and for the texture features that carry much of radiomics' discriminative value, this protocol-induced variation can rival or exceed the biological differences of interest. The same dependence undermines deep learning models, whose performance often falls on data acquired outside the training distribution. Existing responses largely seek to reduce a biomarker's sensitivity to protocol, through harmonization, standardization, or stability-based feature selection. This dissertation takes a complementary approach: rather than reduce protocol sensitivity, it characterizes and predicts it in terms of physical, quantitative image quality metrics. To enable this, an open-source CT texture phantom with 3D-printed modular inserts was designed and manufactured, producing distinct and reproducible radiomic signatures. The phantom was scanned together with an image quality phantom across 170 protocols spanning four scanner models from two vendors, pairing measurements of three-dimensional image quality (MTF, NPS, NEQ, and CNR) with radiomic texture features extracted under identical conditions. Analysis of this dataset established that texture features respond systematically to fundamental image quality characteristics, with relationships largely consistent across different physical textures. Regression models were then developed to predict radiomic feature values from image quality metrics. Finally, the framework was extended to a convolutional neural network classifying lung nodules, trained and evaluated across reconstruction kernels. The change in classifier performance was predictable from a single phantom-derived resolution metric, and a model fit on one scanner retained much of its accuracy when transferred to an unseen scanner, providing initial evidence that protocol-induced performance degradation can be anticipated from phantom measurements alone. By describing protocol sensitivity in vendor-neutral physical terms, this work moves toward defining the range of CT conditions over which a quantitative imaging biomarker can be reliably applied, an imaging inclusion criterion expressed in measurable image quality rather than manufacturer labels, and toward predicting prospectively where such models would fail on out-of-distribution clinical data.