Author: Aditya P. Apte, Joseph O. Deasy, Yusuf Emre Erdi, Anqi Fu, Johannes Hertrich, Andrew Jackson, Usman Mahmood, Jason Ocana, Trahan Sean, Amita Shukla-Dave 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center 🌍
Purpose: Automated AI-based quantitative CT tools hold immense promise for advancing clinical decision-making, yet their reproducibility and generalizability remain vulnerable to variability in imaging protocols, reconstruction parameters, and processing pipelines. Existing CT quality control (QC) phantoms are effective for measuring noise, resolution, and uniformity but lack the anatomical complexity required to validate feature-dependent AI workflows. Here, we assess the feasibility of voxel printing a specialized phantom that is both standardized and clinically relevant, aiming to facilitate harmonization and reproducibility across diverse AI-based CT analysis protocols.
Methods: A thoracic phantom was manufactured using voxel printing (VP), a polyjet 3D printing technology with Radiomatrix material (HU range: -30 to 1000). VP enabled controlled x-ray attenuation through spatial half-toning at sub-voxel resolution (42×84×30 µm). An entire section of a patient's thorax was replicated using ray-tracing based segmentation to capture detailed lung vasculature. Five scans were acquired using a routine CT chest protocol with a 64-slice GE HD750 scanner. Quantitative comparison included multi-scale structural similarity metrics (MS-SSIM) and 104 radiomic features extracted using PyRadiomics.
Results: The phantom demonstrated high MS-SSIM with the patient data (MS-SSIM: 0.98-1.0) and preserved geometric characteristics with shape-based features closely matching patient values (elongation: 0.74 vs 0.71; surface area: 8005mm² vs 7662mm²). While CT numbers showed systematic offset (ROI means: -411 HU vs -47 HU), texture pattern relationships were preserved across multiple matrices, evidenced by similar entropy values (GLCM: 8.83 vs 8.97; GLRLM: 5.34 vs 5.22; GLSZM: 7.15 vs 7.14). Intensity distributions closely paralleled patient patterns, indicating stable relative contrasts despite minor discrepancies.
Conclusion: This work demonstrates a novel technical approach combining voxel printing with radiomatrix material to create reproducible reference phantoms. The preserved relative tissue contrasts and stable feature relationships, despite systematic HU offsets, provide a foundation for harmonization of radiomic and AI-based quantitative CT analysis tools.