Computational Toolkit for Normalizing the Impact of CT Acquisition and Reconstruction on Quantitative Image Features

点击翻译按钮获取中文摘要

Quantitative image features (QIFs) such as radiomic and deep features hold enormous potential to improve the detection, diagnosis, and treatment assessment of a wide range of diseases. Generated from clinically acquired Computed Tomography (CT) scans, QIFs represent small pixel-wise changes that may be early indicators of disease progression. However, detecting these changes is complicated by variations in the way that CT scans are performed, including variations in acquisition and reconstruction parameters. Ensuring reproducible QIFs is a prerequisite for developing machine learning (ML) models that achieve consistent performance across different clinical settings. This project's premise is that QIFs are sensitive to CT parameters such as radiation dose level, slice thickness, reconstruction kernel, and reconstruction method. The combined interactions among these parameters result in unique image conditions, each yielding its own QIF value. Moreover, some clinical tasks and algorithms are more sensitive to differences in QIF values than others. We hypothesize that a systematic, task-dependent framework to normalize scans and mitigate the impact of variability in CT parameters will identify reproducible QIFs and yield more consistent ML models. Three interrelated innovations will be pursued in this work: 1) a novel framework for characterizing the impact of different acquisition and reconstruction parameters on QIFs and ML models using patient scans with known clinical outcomes in multiple domains; 2) a systematic approach for selecting an optimal mitigation technique and evaluating the impact of normalization; and 3) an open-source software toolkit that formalizes the process of CT normalization, addressing real-world use cases developed by academic and industry collaborators. In Aim 1, we will evaluate how multiple CT parameters influence QIF values and model performance. Utilizing metrics of agreement and a heat map-based visualization, we will determine under which image acquisition and reconstruction conditions the QIFs and model performance are consistent. In Aim 2, we will develop and validate a generative adversarial network-based approach to normalization. Our investigation will focus on targeted mitigation of the set of imaging conditions that are most relevant to a clinical task and on the optimization of how these models are trained. In Aim 3, we will engage a spectrum of external stakeholders to guide the development and adoption of a software toolkit called CT-NORM. Three distinct clinical domains will drive our efforts: lung nodule detection (which relies on identifying small regions of high contrast differences to identify nodules), interstitial lung disease quantification (which depends on characterizing texture differences), and ischemic core assessment (which relies on detecting low contrast differences in brain tissue). CT-NORM will provide the scientific community with an approach and a unified toolkit to characterize and mitigate the impact of reconstruction and acquisition parameters on QIFs and ML model performance. By addressing critical sources of variability, we will improve the process of generating QIFs and facilitate the discovery of precise and reproducible imaging phenotypes of disease.