A task-based phantom investigation evaluating computed tomography performance has demonstrated that deep-learning iterative reconstruction combined with high-frequency kernels substantially reduces image noise while preserving spatial resolution, but introduces notable contrast penalties for subtle structures at low radiation doses. Conducted across standardized chest and lumbar spine protocols, the technical assessment evaluated General Electric Healthcare’s deep-learning iterative reconstruction algorithm, designated DLIR, against conventional adaptive statistical iterative reconstruction, designated ASIR-V. While DLIR configured at its highest strength yielded superior overall lesion detectability under routine clinical dose conditions, its aggressive noise filtering caused pronounced contrast reduction and inferior detectability for small, low-concentration calcium-based targets when radiation exposure was substantially lowered. These experimental findings establish important operational boundaries for artificial intelligence algorithms applied to edge-enhancing and bone-specific convolution kernels, demonstrating that deep neural networks do not uniformly outperform established iterative methods across every diagnostic task and dose level.
Over the past decade, iterative reconstruction algorithms have served as an essential pillar of medical computed tomography by replacing traditional filtered back projection to reduce quantum noise and minimize patient radiation exposure. Iterative processing enabled clinicians to achieve meaningful dose reductions across routine computed tomography protocols while preserving diagnostic image quality. Nevertheless, high-strength iterative reconstruction techniques frequently exhibit non-linear noise behavior that manifests as an unnatural, blotchy, or plastic image texture. Radiologists have widely noted that this artificial appearance can degrade subjective viewing confidence and alter the visual presentation of fine anatomical boundaries. To overcome these aesthetic and diagnostic shortcomings, medical device manufacturers developed deep-learning image reconstruction systems designed to denoise volumetric scans while preserving a more traditional, natural grain texture reminiscent of filtered back projection. When General Electric introduced its deep-learning iterative reconstruction platform in 2020, the neural network was initially restricted to smooth soft-tissue reconstruction kernels, leaving high-resolution clinical imaging dependent on conventional iterative filters.
High-frequency reconstruction kernels, such as edge-enhancing thoracic filters and sharp musculoskeletal kernels, are routine fixtures in clinical protocols that demand elevated spatial resolution, including pulmonary parenchymal evaluation and detailed bone examinations. Unlike smoothing kernels that suppress fine fluctuations, high-frequency algorithms intentionally amplify high spatial frequencies to sharpen the boundaries of small blood vessels, bronchovascular bundles, and delicate trabecular bone architecture. This intentional edge enhancement inevitably multiplies image noise, generating a coarse and grainy visual background that challenges reconstruction engines. Distinguishing legitimate microstructural signal from background noise becomes exceptionally demanding when small pathological lesions present with low intrinsic contrast. Although preliminary clinical work had suggested that extending deep-learning iterative reconstruction to high-frequency kernels improved the delineation of pancreatic tumor margins, objective task-based evaluations of this technological union had never been documented in medical physics literature. The current study aimed to fill that evidential gap by conducting the first comprehensive phantom assessment of spatial resolution, noise power spectra, lesion contrast, and mathematical observer detectability across clinical dose ranges.
To execute this rigorous technical evaluation, the research team employed three dedicated physical imaging phantoms scanned on a commercial GE Revolution Apex Elite computed tomography scanner. The acquisition parameters were held constant across all testing sequences at a tube potential of 120 peak kilovoltage, a helical pitch of 0.992, a gantry rotation time of 0.35 seconds, a detector collimation of 64 by 0.625 millimeters yielding a total beam width of 40 millimeters, an axial slice thickness of 1.2 millimeters, and a reconstruction increment of 0.625 millimeters. Two distinct high-frequency convolution kernels were examined: the edge-enhancing Lung kernel for thoracic imaging and the high-resolution Bone-plus kernel for lumbar spine imaging. Tube currents were modulated systematically to simulate both standard clinical protocols and aggressively lowered exposure levels. In-plane spatial resolution, longitudinal spatial resolution, and three-dimensional noise power spectra were quantified using a water-filled cubic test object measuring 160 millimeters on each side with rounded corners to avert streak artifacts, housing a removable central polymethyl methacrylate cylinder measuring 70 millimeters in diameter and height.
Lesion contrast and task-based detection performance were appraised using two semi-anthropomorphic elliptical phantoms representing thoracic and abdominal anatomies, manufactured by QRM. An insert module containing nine cylindrical hydroxyapatite calcifications embedded in soft-tissue-equivalent background material was positioned inside both anthropomorphic bodies. These simulated calcium lesions incorporated three distinct physical diameters of 1, 3, and 5 millimeters paired with three material concentrations of 200, 400, and 800 milligrams of hydroxyapatite per cubic centimeter, each cylinder maintaining a height matching its diameter. The cubic phantom was scanned five times per dose level at 0.8, 3.0, and 7.0 milligrays for the chest protocol, and at 3.0, 7.0, and 15.0 milligrays for the spine protocol. The thoracic phantom underwent ten independent scans at six dose levels ranging from 0.2 to 7.0 milligrays, while the abdominal phantom was scanned ten times across six dose levels ranging from 0.5 to 15.0 milligrays. Reconstructions were executed across five algorithmic settings: baseline ASIR-V at zero percent representing raw filtered back projection, ASIR-V at eighty percent strength, and deep-learning iterative reconstruction configured at Low, Medium, and High strengths.
Assessment of in-plane spatial resolution relied on computing the target transfer function along the circular boundary between the central polymethyl methacrylate rod and surrounding water, which produced a nominal baseline contrast of 120 Hounsfield units. Regions of interest measuring 100 by 100 millimeters were extracted from 80 axial slices, yielding a contrast-to-noise ratio of 57 even in the most unfavorable scanning condition, easily exceeding the recommended minimum threshold of 15 required to restrict computational uncertainty below 10 percent. Radial edge spread functions were sampled in 5-degree increments over 10-degree angular apertures and Fourier-transformed to generate normalized in-plane target transfer functions. Across both kernels and all exposure levels, in-plane resolution exhibited pronounced edge enhancement that peaked higher for the Lung kernel than for the Bone-plus kernel. While ASIR-V at eighty percent strength suffered noticeable degradation in spatial resolution and a downward peak frequency shift as radiation dose decreased, deep-learning reconstructions maintained rock-solid in-plane resolution across all three strength levels, performing slightly better than raw filtered back projection. Conversely, longitudinal spatial resolution measured along the scanner z-axis proved entirely independent of reconstruction algorithms and kernels, governed solely by slice thickness and increment.
Three-dimensional noise power spectra were calculated from a 100 by 100 by 100 millimeter volume of interest placed within the uniform water section of the cubic phantom, with no artificial windowing or detrending applied. Noise magnitude decreased with increasing radiation dose and was consistently higher when using the Lung kernel compared to the Bone-plus kernel. Deep-learning iterative reconstruction delivered remarkable noise suppression, slashing noise magnitude by factors between 1.9 and 2.8 for the Lung kernel and between 2.0 and 2.8 for the Bone-plus kernel relative to raw baseline processing. While ASIR-V at eighty percent compressed the noise spectrum toward lower spatial frequencies—reducing average frequency values to between 0.382 and 0.390 inverse millimeters for Lung and 0.613 to 0.622 inverse millimeters for Bone-plus, generating a conspicuously blurred texture—DLIR preserved higher spatial frequencies ranging from 0.423 to 0.463 inverse millimeters and 0.659 to 0.698 inverse millimeters, successfully maintaining the crisp, natural appearance characteristic of unfiltered scans.
Quantitative evaluation of lesion contrast revealed an important performance trade-off. Automated sampling on 3-millimeter and 5-millimeter calcifications showed that while ASIR-V preserved contrast near baseline across all exposures, DLIR exhibited progressive contrast loss with increasing strength. This penalty worsened severely at low radiation doses for small, low-concentration targets. In the lumbar spine protocol at 0.5 milligrays, DLIR-High caused a 70 percent contrast reduction for the 3-millimeter 200 milligram-per-cubic-centimeter lesion, and an 82 percent loss at 0.2 milligrays in the chest protocol. In contrast, at standard clinical doses of 7 and 15 milligrays, contrast differences remained minimal at under 5 to 11 percent.
Lesion detectability was quantified using a non-prewhitening model observer with an eye filter, modeling human visual performance. At standard diagnostic doses, DLIR-High achieved superior detectability across all lesion types, outperforming ASIR-V at eighty percent by 53 percent and 51 percent at 3 and 7 milligrays in the chest protocol. However, at ultra-low dose levels, severe contrast loss eroded this detection advantage. For the 3-millimeter calcification at 200 milligrams per cubic centimeter, DLIR-High detectability fell 42 percent below ASIR-V eighty percent at 0.8 milligrays in the chest, and dropped 45 percent lower at 3 milligrays and 20 percent lower at 7 milligrays in the spine protocol.
The authors attributed this contrast attenuation to deep-learning behavior under low contrast-to-noise ratios. When photon starvation obscures subtle features, the neural network cannot reliably distinguish weak structural signals from background noise, resulting in over-smoothing of delicate abnormalities.
Because this evaluation relied on homogeneous physical phantoms on a single scanner, it cannot fully capture anatomical complexity, tissue heterogeneity, patient motion, or varied patient body habitus. Additionally, proprietary algorithms from other manufacturers require independent task-based validation. The authors concluded that while deep learning with high-frequency kernels provides transformative noise reduction at routine doses, human reader clinical studies remain essential to confirm diagnostic efficacy across challenging low-dose diagnostic tasks.
Subject of Research: Task-based image quality evaluation of deep-learning iterative reconstruction with high-frequency CT kernels
Article Title: Evaluation of deep-learning iterative reconstruction combined with high-frequency kernels in CT: a task-based image quality study
Article References: Viry, A., Monnin, P., Pozzessere, C., & Racine, D. (2026). Evaluation of deep-learning iterative reconstruction combined with high-frequency kernels in CT: a task-based image quality study. European Radiology Experimental, 10(1), Article 128. https://doi.org/10.1186/s41747-026-00801-3
Image Credits: AI Generated
DOI: 10.1186/s41747-026-00801-3
Keywords: computed tomography, deep-learning iterative reconstruction, high-frequency kernels, image quality, lesion detectability, noise power spectrum, target transfer function, radiation dose, ASIR-V, anthropomorphic phantoms, non-prewhitening model observer, chest CT, lumbar spine CT, spatial resolution, contrast-to-noise ratio, hydroxyapatite calcifications, filtered back projection
News Source: Ophelia Keating. (November 12, 2026). Deep-Learning CT Reconstruction Cuts Noise but Suppresses Subtle Low-Dose Lesions. Scienmag.



