Deep Learning Reconstruction versus Hybrid Iterative Reconstruction for Spectral CT: A Multi-Institutional Image Quality Assessment
- 1Hospital of the University of Pennsylvania
Abstract
PurposeTo compare image quality between deep learning reconstruction and hybrid iterative reconstruction for conventional and virtual monoenergetic images derived from the same spectral CT examinations.
Methods and MaterialsThis retrospective, post hoc secondary analysis of a multi-institutional, controlled, blinded reader study evaluated head, chest, cardiac, and abdomen/pelvis CT examinations acquired using dual-layer spectral CT. The parent study (clinicaltrials.gov identifier NCT07108205) included 147 examinations across a range of acquisition and reconstruction parameters. Raw data were reconstructed using hybrid iterative reconstruction (HIR; iDose4, Philips Healthcare) and spectral deep learning reconstruction (DLR; Spectral Precise Image, Philips Healthcare). Conventional images and corresponding virtual monoenergetic images (VMI) were generated with both methods from identical raw data. Randomized image pairs were independently evaluated by two blinded board-certified radiologists or cardiologists, and image quality was scored on a 5-point Likert scale. A post hoc analysis was performed on previously collected scores from a subset of examinations acquired at standard resolution and reconstructed with thin sections and a soft tissue kernel to focus comparison on HIR versus DLR. Least squares means were estimated using a model accounting for multiple readers per examination. Noninferiority of DLR to HIR was tested; when established, superiority was tested using the same model.
Results59 CT examinations from 56 patients met technical criteria for this analysis. DLR was noninferior and superior to HIR for head, abdomen, and cardiac examinations across conventional images and VMI, with p values ranging from <.0001 to .0058. DLR was noninferior but not superior for conventional chest images (p = .1430). Noninferiority was not established for chest VMI, although mean scores numerically favored DLR. This may reflect a ceiling effect from high HIR scores.
ConclusionSpectral DLR produced image quality that matched or exceeded HIR across most anatomic regions and image types, including conventional and virtual monoenergetic images.
Related
- Deep Learning–Based Reconstruction in Ultra-Low-Dose CT of the Ankle and Foot: A Comparative Study with Hybrid Iterative Reconstruction
- Quantification of plaque and detection of high-risk plaque characteristics based on coronary computed tomography angiography: comparison between superresolution deep learning reconstruction and hybrid iterative reconstruction
- Multidimensional institutional learning curve of ultrasound-guided percutaneous transthoracic lung biopsy: a cumulative sum analysis with case-mix assessment
- Comparing a Dedicated Breast Spectral CT Protocol With Standard MonoE40 Reconstruction for Breast Cancer Characterization and Opportunistic Assessment
- Evaluation of accelerated whole-body diffusion weighted imaging with deep learning reconstruction in patients with metastatic prostate cancer: assessment of image quality and ADC estimates.
BibTeX
@misc{nazarian2026deep,
title = {{Deep Learning Reconstruction versus Hybrid Iterative Reconstruction for Spectral CT: A Multi-Institutional Image Quality Assessment}},
author = {M. Nazarian and A. E. Perkins and E. Y. Wong and P. C. Koopmans and M. B. Stromza and L. Roshkovan and M. Mathew and J. Dorff and J. M. Levsky and M. S. Lazarus and P. B. Noël and L. Slipczuk and A. Dhanaliwala},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.10.26362104},
url = {https://www.medrxiv.org/content/10.64898/2026.09.10.26362104v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Deep Learning Reconstruction versus Hybrid Iterative Reconstruction for Spectral CT: A Multi-Institutional Image Quality Assessment
AU - M. Nazarian
AU - A. E. Perkins
AU - E. Y. Wong
AU - P. C. Koopmans
AU - M. B. Stromza
AU - L. Roshkovan
AU - M. Mathew
AU - J. Dorff
AU - J. M. Levsky
AU - M. S. Lazarus
AU - P. B. Noël
AU - L. Slipczuk
AU - A. Dhanaliwala
PY - 2026
DA - 2026/09/10
PB - medRxiv
DO - 10.64898/2026.09.10.26362104
UR - https://www.medrxiv.org/content/10.64898/2026.09.10.26362104v1
AB - PurposeTo compare image quality between deep learning reconstruction and hybrid iterative reconstruction for conventional and virtual monoenergetic images derived from the same spectral CT examinations. Methods and MaterialsThis retrospective, post hoc secondary analysis of a multi-institutional, controlled, blinded reader study evaluated head, chest, cardiac, and abdomen/pelvis CT examinations acquired using dual-layer spectral CT. The parent study (clinicaltrials.gov identifier NCT07108205) included 147 examinations across a range of acquisition and reconstruction parameters. Raw data were reconstructed using hybrid iterative reconstruction (HIR; iDose4, Philips Healthcare) and spectral deep learning reconstruction (DLR; Spectral Precise Image, Philips Healthcare). Conventional images and corresponding virtual monoenergetic images (VMI) were generated with both methods from identical raw data. Randomized image pairs were independently evaluated by two blinded board-certified radiologists or cardiologists, and image quality was scored on a 5-point Likert scale. A post hoc analysis was performed on previously collected scores from a subset of examinations acquired at standard resolution and reconstructed with thin sections and a soft tissue kernel to focus comparison on HIR versus DLR. Least squares means were estimated using a model accounting for multiple readers per examination. Noninferiority of DLR to HIR was tested; when established, superiority was tested using the same model. Results59 CT examinations from 56 patients met technical criteria for this analysis. DLR was noninferior and superior to HIR for head, abdomen, and cardiac examinations across conventional images and VMI, with p values ranging from <.0001 to .0058. DLR was noninferior but not superior for conventional chest images (p = .1430). Noninferiority was not established for chest VMI, although mean scores numerically favored DLR. This may reflect a ceiling effect from high HIR scores. ConclusionSpectral DLR produced image quality that matched or exceeded HIR across most anatomic regions and image types, including conventional and virtual monoenergetic images.
ER -