Author: John M. Boone, Andrew M. Hernandez, Paul E. Kinahan, Michael F. McNitt-Gray, Jeffrey H. Siewerdsen, Ali Uneri 👨🔬
Affiliation: University of California, Johns Hopkins Univ, UT MD Anderson Cancer Center, David Geffen School of Medicine at UCLA, University of Washington, UC Davis Health 🌍
Purpose: Measuring image quality (IQ) in large clinical databases, such as the Medical Imaging and Data Resource Center (MIDRC), is challenging due to the inherent complexity of image content and the intricate methodologies required for IQ evaluation. While DICOM metadata provides key scanner and protocol details, it falls short in providing metrics that directly relate to IQ. The goal of this study is to provide the framework for MIDRC users to query the database and select custom-tailed cohorts of images using IQ descriptors.
Methods: The approach leverages a proxy dataset of phantom images to map measured IQ metrics (e.g., 3D MTF and 3D NPS) onto clinical chest CT images from the MIDRC database. The multi-institutional dataset comprises >250 CT series of the Corgi phantom (6 institutions, 4 vendors, 11 models). Images were acquired using institution-specific non-contrast chest CT protocols and were matched to clinical images using select DICOM attributes, including manufacturer, scanner model, tube voltage, tube current, exposure time, convolutional kernel, and voxel spacing.
Results: Of the 30,000 chest CT images in the MIDRC database, >50% were represented by scanner models included in the phantom dataset. Preliminary efforts using strict matching criteria (±10 kV, ±20% exposure, ±0.1 mm voxel spacing) yielded ~1,000 matches. Cohorts selected based on predicted CNR and 3D MTF metrics exhibited qualitatively distinct features. Validation using radiomic features (e.g., gray level run length matrix) confirmed the differences, such as significantly increased run variance for images with higher CNR, and increased short-run emphasis for images with higher MTF.
Conclusion: This study establishes a framework for assigning IQ descriptors to clinical images in the MIDRC database. The capability to select cohorts using well-defined IQ metrics facilitates systematic testing of image-based algorithms and supports evaluations of their generalization across images with diverse qualities.