Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session E

Poster R6 - Effect of Reconstruction Kernel on Voxel-Based Image Analysis of Chest Computed Tomography Scan

Saturday, October 26, 2024

10:00 AM - 11:00 AM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Olivia A. Colavito (she/her/hers)

Student
Rutgers University
Aberdeen, New Jersey, United States

Co-Author(s)

GP

Gregory Pappas, PhD

Research Scientist
Department of Veterans Affairs NJ War Related Illness & Injury Study Center (WRIISC), United States
TA

Thomas Abitante, PhD (he/him/his)

Postdoctoral Researcher
Department of Veterans Affairs NJ War Related Illness & Injury Study Center (WRIISC), United States
BS

Bishoy Samy, BS, MS

Biomedical Engineer
Department of Veterans Affairs NJ War Related Illness & Injury Study Center (WRIISC), United States

Primary Investigator(s)

MF

Michael Falvo, PhD

Co-Director of Airborne Hazards and Burn Pits Center of Excellence (AHBPCE)
Department of Veterans Affairs NJ War Related Illness & Injury Study Center (WRIISC), United States

Introduction: High-resolution computed tomography (HRCT) of the chest is clinically indicated for the diagnosis of lung disease. Quantitative analysis of HRCT – via parametric response mapping (PRM) – provides a voxel-by-voxel analysis of lung tissue, allowing for characterization of normal, functional low-density, or persistent low-density features. Functional and persistent low-density areas are indicative of small airways disease and emphysema, respectively. It is possible for a voxel to be unclassified, with no clear identification. Raw CT files are reconstructed to create lung images for analysis using kernels that influence resolution (e.g., sharpness, noise). Kernel types vary across scanners and manufacturers and are known to influence PRM parameters. However, the impact of kernel type on PRM parameters is not well studied and overlooked among clinicians and researchers. The purpose of our study is to assess the impact of two kernel types (sharp, smooth) on PRM algorithm performance.

Materials and

Methods: 76 chest HRCTs (paired inspiratory and expiratory scans) were obtained from non-smoking veterans participating in a respiratory health research study. HRCTs were acquired randomly on one of two scanners and their corresponding kernels (Toshiba, FC09 sharp kernel; Phillips, B smooth kernel). Images underwent an anonymization process to remove patient identifiers using open-source (MicroDicom, DicomEditor) and custom (Matlab) software. PRM analysis was performed on anonymized images using a lung density analysis software package (Imbio, LLC). PRM analysis yielded percentage of voxels that were classified (normal, functional, persistent low-density) or unclassified. We compared the percentage of unclassified voxels between kernels via t-test (2-tailed) and effect size (Cohen’s d).

Results, Conclusions, and Discussions: HRCT images were acquired using a Toshiba or Phillips CT scanner, with either a sharp (n = 46) or smooth (n = 30) kernel, respectively. Participant demographics were similar between scanner/kernels (sharp vs. smooth; 42.0 vs. 45.9 years of age, 88.0% male vs. 86.1% male, 30.3 vs. 30.4 kg/m^2). A greater percentage of unclassified voxels was observed among scans reconstructed with a sharp kernel (20.7 ± 6.0% vs. 0.6 ± 0.7%, p < 0.001, d = 4.23).

In our sample, we found that approximately 20% of voxels on average were unable to be classified by PRM analysis when reconstructed using a sharp kernel, but fewer than 1% were unclassified when using a smooth kernel. Our findings emphasize the significant influence of kernel type on PRM analysis. Clinicians and researchers should be cautious when interpreting studies that rely on images reconstructed with sharp kernels. Efforts to improve comparability across studies with various kernel types using normalization techniques are actively being studied and should continue to be rigorously pursued for the analysis of retrospectively collected data. For prospective CT acquisition, scanning protocols should be thoughtfully considered to allow for standardized reconstructions ideal for the viewing radiologist (e.g., sharper kernels), as well as for voxel-based analysis (e.g., smoother kernels). Our observed differences contribute to the slow adoption of quantitative CT analysis integrated within clinical workflows. For example, clinicians may be reluctant to implement quantitative CT analysis within their diagnostic approach due to the unknown effects of image processing techniques, such as kernel reconstruction. Further, voxel-based analysis techniques often utilize proprietary algorithms and artificial intelligence-based techniques that lack transparency and contribute to mistrust. Still, the considerable potential of voxel-based analysis for making earlier diagnoses and monitoring treatment outcomes is too important to ignore. Until then, clinicians and researchers must be aware of the effect of kernel type on voxel-based analysis software programs and their corresponding results.

Acknowledgements (Optional):