PROJECT SUMMARY / ABSTRACT Radiomics, or imaging biomarkers, are an active area of research and development that is increasing in breadth with more widespread access to large, patient image databases. Radiomics models have been applied in a wide range of diagnostics, classification tasks, and disease scoring; with advantages for efficient radiology workflow, reducing errors and highlighting important features, and providing additional information in challenging diagnostic cases. Accuracy of radiomics is dependent on a number of factors. The variability associated with the imaging chain including the particular imaging device/vendor, acquisition protocol, data processing, etc. is undesirable and can have a dramatic effect on a radiomics model’s performance. Successful radiomics models generally require careful data curation and standardization of protocols – often preventing successful or efficient modeling in large aggregations of patient data across institutions, vendors, etc. Moreover, even with careful attention to protocol, many imaging devices, like x-ray computed-tomography CT have patient- and scan-specific image properties that continue to add undesirable variability to a radiomics computation. In this work, we propose a framework for end-to-end modeling of a CT imaging system – integrating radiomics calculations as an explicit stage and imaging system output. This kind of rigorous modeling extends previous efforts to under- stand and control the performance of imaging systems. In this context, the proposed mathematical framework provides not only a mechanism for prediction of radiomics values based on the various system depend- ences that degrade their accuracy; but also informs recovery approaches to estimate the underlying “true” radiomics based on the underlying biology uncorrupted by the particular image properties (noise/resolution) of the patient image. We hypothesize that this new paradigm for radiomics computation will both standardize met- rics and improve quantitation. We will test these hypotheses and apply standardization methods to radiomics for interstitial lung disease (ILD, an application where lung textures provide substantial diagnostic information about the disease) through the following specific aims: Aim 1: Develop a mathematical framework for radiomics standardization, wherein both predictive “forward” models and “inverse” recovery models for ILD radiomics will be developed, characterized, and evaluated. Aim 2: Apply and validate prediction and standardization framework in physical systems using custom phantoms with lung textures and including a series of investiga- tions on well-characterized CT benches and CT scanners from all major vendors. Aim 3: Investigate the impact of standardization on radiomics modeling performance in clinical CT data. A multi-site study will establish the performance of standardized radiomics using the proposed framework in radiomics models for both regional and whole lung characterization. Successful completion of these aims will establish a new paradigm for stand- ardized radiomics computation that is applied and validate in multi-site data. This opens the doors to larger, more diverse imaging datasets and the potential for more efficient recovery of subtle imaging biomarkers.

PROJECT NARRATIVE Radiomics models, or imaging biomarkers, have been proposed to allow automatic computer-aided diagnosis, classification, and scoring of various diseases based on imaging study data. Radiomics research and development is a rapidly growing field with the increasing availability of large imaging databases and access to sophisticated software and modeling tools; however, successful radiomics models typically require careful data curation and understanding to limit variability associated with the imaging chain (in contrast to the desired variability in the underlying patient anatomy which contains the target biomarkers). This work seeks to provide a rigorous mathematical framework with models for predicting the undesirable variability associated with the imaging device, acquisition protocol, and data processing; as well as strategies for standardization across those dependencies to rigorously estimate the underlying imaging biomarkers, thus permitting high-performance radiomics model development and application across a broad range of data sources.