An AI/ML-ready Dataset for Investigating the Effect of Variations in CT Acquisition and Reconstruction

定量图像特征（qif），如放射学和深度特征，在改善图像质量方面具有巨大的潜力

Quantitative image features (QIFs) such as radiomic and deep features hold enormous potential to improve the detection, diagnosis, and treatment assessment of various diseases. When extracting QIFs from computed tomography (CT) scans, computed values can vary based on differences in CT acquisition and reconstruction parameters, including radiation dose level, slice thickness, reconstruction kernel, and reconstruction method. The performance of artificial intelligence (AI) and machine learning (ML) models depends on the diversity of data on which the model was trained. Previous studies have shown the negative impact that differences in CT acquisition and reconstruction have on the reproducibility of radiomic feature values and the performance of AI/ML models. However, there is a dearth of real-world datasets that enable AI/ML developers and researchers can easily leverage to train and validate models that are robust to these differences. The objective of this supplement is to improve the AI/ML-readiness of real-world patient CT datasets, facilitating investigations into characterizing and mitigating the effect of variations in CT acquisition and reconstruction parameters. This project builds upon our parent R01 project (R01 EB031993, Computational Toolkit for Normalizing the Impact of CT Acquisition and Reconstruction on Quantitative Image Features), which aims to understand the effect of these variations on downstream AI/ML models and clinical tasks (e.g., nodule detection, stroke characterization) and develop effective methods for image harmonization. This project will bring together expertise in informatics, medical physics, and data/model sharing standards. In Aim 1, we will release an AI/ML-ready CT dataset of 200 chest CT scans of patients who underwent lung cancer screening and 100 non-contrast head CTs of patients with suspected stroke. Each scan will be reconstructed by varying dose, slice thickness, and kernel, resulting in over 30 different versions of the same scan. Scans will also be annotated (e.g., outlined nodule boundaries) and linked with clinical information (e.g., nodule characteristics, pathology-confirmed lung cancer diagnosis). Following FAIR principles, clinical data, scans, and annotations will be released using established common data elements and standards such as DICOM segmentation objects. In Aim 2, we will demonstrate the utility of this dataset as a benchmark for assessing the reliability and robustness of AI/ML algorithms. We will use the benchmark CT dataset to evaluate the performance of publicly available algorithms for lung nodule detection and characterization and ischemic volume estimation. We will assess the robustness of these algorithms’ performance using metrics such as sensitivity and false positives/scan (nodule detection), area under the receiver operating characteristic curve (nodule classification), and mean absolute error (stroke quantification) across different scans. Successful completion of this project will result in a unique dataset that would double the available real-world patient data that can be used to improve AI/ML algorithms related to image reconstruction, restoration/harmonization, and downstream tasks.