Phantom-based performance comparison of two commercial deep learning CT reconstruction algorithms with super- and normal-resolution settings

Joël Greffier; Catherine Roy; Djamel Dabli; Jean-Paul Beregi; Maxime Pastor

doi:10.1186/s41747-025-00670-2

Phantom-based performance comparison of two commercial deep learning CT reconstruction algorithms with super- and normal-resolution settings

Eur Radiol Exp. 2026 Jan 26;10(1):9. doi: 10.1186/s41747-025-00670-2.

Authors

Joël Greffier¹, Catherine Roy², Djamel Dabli³, Jean-Paul Beregi⁴, Maxime Pastor⁴

Affiliations

¹ IMAGINE UR UM 103, Montpellier University, Department of Medical Imaging, Nîmes University Hospital, Nîmes, France. joel.greffier@chu-nimes.fr.
² Diagnostic Imagery Department, Nouvel Hôpital Civil (NHC), Hôpitaux Universitaires de Strasbourg, Strasbourg, France.
³ IMAGINE UR UM 103, Montpellier University, Department of Medical Imaging, Nîmes University Hospital, Nîmes, France. djamel.dabli@chu-nimes.fr.
⁴ IMAGINE UR UM 103, Montpellier University, Department of Medical Imaging, Nîmes University Hospital, Nîmes, France.

Abstract

Objective: We compared a super-resolution deep learning image reconstruction (SR-DLR) algorithm with a normal-resolution (NR)-DLR algorithm according to radiation dose for abdominal computed tomography (CT).

Materials and methods: An image-quality phantom was scanned with an energy-integrating detectors CT unit at three volume CT dose index radiation dose levels (12.7, 5.9, and 3 mGy). Images were reconstructed using a 1,024² matrix for SR-DLR and a 512² matrix for NR-DLR, for three DLR levels (level-1, level-2, and level-3). Noise power spectrum (NPS) and task-based transfer function (TTF) for iodine and Solid Water^® inserts were computed; TTF values at 50% (f₅₀, mm^-1) were used to quantify spatial resolution. The detectability index (d') was computed for two simulated lesions.

Results: Noise magnitude values were lower with SR-DLR than with NR-DLR for level-2 (-27.6 ± 3.8%) and level-3 (-43.5 ± 1.4%), the opposite for level-1. Average NPS spatial frequency was higher with SR-DLR than with NR-DLR for all radiation dose levels for level-1 (55.9 ± 16.7%) and level-2 (20.1 ± 13.9%) and the opposite for level-3, except at 12.7 mGy. For both inserts, f₅₀ was higher with SR-DLR than with NR-DLR at each radiation dose and DLR level. For simulated lesions and all DLR levels, d' values were higher with SR-DLR than with NR-DLR (level-1, 6.0 ± 2.0%; level-2, 45.7 ± 5.0%; level-3, 75.2 ± 7.3%).

Conclusion: Compared to NR-DLR, SR-DLR improved spatial resolution and detectability of simulated abdominal lesions; image noise was reduced with SR-DLR only for level-2 and level-3, while image texture was better for level-1 and level-2.

Relevance statement: Super-resolution DLR with a 1,024² matrix size improved spatial resolution and detectability of simulated abdominal lesions compared to normal-resolution DLR. Validation in clinical settings is necessary before translation into routine practice.

Key points: The performance of a new deep learning super-resolution image reconstruction algorithm (SR-DLR) was compared to a normal-resolution (NR)-DLR algorithm using an image-quality phantom for an abdominal energy-integrating detector CT protocol. SR-DLR with a 1,024² matrix improved spatial resolution and detectability of simulated abdominal lesions compared to NR-DLR with a 512² matrix. Using SR-DLR, therefore, presents numerous prospects for improving abdominal CT images and a high potential for reducing the radiation doses.

Keywords: Artificial intelligence; Deep learning; Image enhancement; Image processing (computer-assisted); Multidetector computed tomography.

Publication types

Comparative Study

MeSH terms

Algorithms*
Deep Learning*
Humans
Image Processing, Computer-Assisted* / methods
Phantoms, Imaging*
Radiation Dosage
Radiographic Image Interpretation, Computer-Assisted* / methods
Tomography, X-Ray Computed* / methods