Citations

Abstract Objective. Four-dimensional CT (4DCT) enables visualization of dynamic wrist joint pathology but may be compromised by motion artifacts. Deep learning (DL)-assisted motion correction has the potential to enhance wrist 4DCT quality, but model development is hindered by scarce training data. To address this, we developed a model-based motion simulation framework that generated realistic 4DCT data with paired static ground truth images to demonstrate wrist motion correction using a physics-informed DL model trained with simulated data. Approach. Wrist motion was simulated by synchronizing rotational and translational movements of carpal bones and connective tissues. Elastic deformation guided by analytically-derived motion vector fields was applied to static wrist scans at neutral positions to create artifact-free images at multiple angular positions. Virtual motion-corrupted 4DCT images were produced via forward projection and reconstruction with synchronized gantry and wrist motion. Experimental data from 4DCT cadaver scans were collected to validate simulation quality. Low-intensity region score (LIRS), positivity, and entropy were used to assess simulated motion artifacts. A physics-informed motion correction network was adapted exclusively using simulated training data and subsequently deployed on simulated, real cadaveric, and in vivo 4DCT scans. LIRS and structural similarity index (SSIM) were used to evaluate motion correction quality. Wilcoxon signed-rank tests were conducted ( α = 0.05). Results. The simulation method created realistic motion artifacts compared to real scan reference ( p > 0.05). Motion artifacts were effectively mitigated across testing cases demonstrating generalizability. In simulated scans, LIRS improved from 0.55 ± 0.12–0.75 ± 0.04 and SSIM from 0.87 ± 0.08–0.95 ± 0.03 at bone regions (each p < 0.01). In experimental cadaveric scans, LIRS improved from 0.05 ± 0.01–0.31 ± 0.17 in marker regions and from 0.58 ± 0.19–0.77 ± 0.14 in bone regions (each p < 0.01). Significance. Our simulation method facilitates training DL-assisted motion correction algorithms by providing high-quality training data, effectively supporting 4DCT image quality enhancement.

Abstract Background Radiomics extracts quantitative imaging features from computed tomography (CT) data for clinical decision‐making. However, variations in acquisition parameters—particularly x‐ray tube voltage (kV)—introduce non‐biological variability in attenuation values, limiting the reproducibility of radiomic features across scanners, protocols, and institutions. Purpose To develop and evaluate a CT dAta harmoNiZAtion framework based on deep learNed vIrTual monoEnergetic imaging (TANZANITE), which leverages the keV flexibility of virtual monoenergetic images (VMIs) to enable cross‐kV scan translation and radiomics harmonization. Methods TANZANITE is a model hub consisting of multiple pre‐trained convolutional neural networks (CNNs), each designed to translate VMIs from 1 keV level to another. Phantom‐based calibration was first used to determine energy‐equivalent (Eff_E) keV levels corresponding to each tube potential (e.g., Eff_E(A) keV for source kV and Eff_E(B) keV for target kV). A CNN trained using 69 120 patches from seven patient cases to map VMIs from Eff_E(A) to Eff_E(B) was selected from the TANZANITE hub and applied directly to clinical CT images acquired at the source kV. This harmonized the images to match the attenuation characteristics of the target kV setting. Evaluation was conducted on independent dual‐energy CT datasets acquired at 100/Sn150 kV. Regions of interest (ROIs) were placed in the kidney, liver, and spine to assess CT number consistency and radiomic feature reproducibility. The concordance correlation coefficient (CCC) was calculated across 93 non‐shape radiomic features. Results After TANZANITE processing with 100 kV images, CT numbers in evaluated organs closely matched the Sn150 kV reference values in four testing patient cases. For example, mean kidney CT numbers changed from 320 HU (100 kV) to 156 HU (TANZANITE), approximating the Sn150 kV value of 160 HU. Similar changes were observed in the liver (157–105 HU vs. 104 HU reference) and spine (45–22 HU vs. 19 HU reference). Radiomic reproducibility improved substantially across organs: mean CCC increased from 0.590 to 0.995 in the liver, 0.300 to 0.970 in the kidney, and 0.630 to 0.968 in the spine. Post‐TANZANITE, over 98% of features exceeded the stability threshold (CCC ≥ 0.900) in all three representative organs. Conclusion TANZANITE provides a flexible, image‐domain harmonization framework by learning the keV‐to‐keV translation in the VMI domain and applying pre‐trained CNNs to clinical kV images. It improves CT number consistency and organ‐specific radiomic reproducibility without requiring raw projection data or scanner‐specific training. This approach supports consistent quantitative imaging across multiple‐kV acquisition protocols, enhancing radiomics reliability in clinical settings.

A deep learning–based neural network enabled virtual monoenergetic imaging at energy-integrating detector CT, reduced blooming artifacts, and improved stenosis assessment in heavily calcified coronary plaques, with performance comparable to ultrahigh-resolution photon-counting detector CT.

Abstract Background Coronary computed tomography angiography (cCTA) is a widely used noninvasive diagnostic exam to assess patients for coronary artery disease (CAD). However, the spatial resolution of most CT scanners is limited due to the use of energy‐integrating detectors (EIDs). Purpose To develop a convolutional neural network ( I mproved LUMEN visualization through A rtificial super‐resolu T ion imag E s (ILUMENATE)) informed by photon‐counting‐detector (PCD)‐CT to improve EID‐CT image resolution and determine its impact on cCTA. Materials and methods With IRB approval, 30 patients undergoing clinically indicated cCTA were scanned with EID‐CT (SOMATOM Force, Siemens Healthineers, Forchheim, Germany) and subsequently with ultra‐high‐resolution (UHR) PCD‐CT (NAEOTOM Alpha, Siemens Healthineers) on the same day. ILUMENATE was trained on eight patient PCD‐CT datasets (67,890 patch pairs with 90% for training (61,101), 10% reserved for validation (6,789)) and applied to 22 unseen EID‐CT cases. Spatial resolution was evaluated using line profiles and percent diameter stenosis quantified with a severity score assigned. Two experienced radiologists, blinded to image type, selected preferred series and scored images for overall quality, sharpness, and noise comparing original EID‐CT and ILUMENATE output. Results Visual assessment and line profiles showed substantial resolution improvement with ILUMENATE. Percent diameter stenosis was significantly reduced (mean ± standard deviation: 4.42% ± 4.82%) using ILUMENATE ( p < 0.001) with nine lesions shifting down in severity score. Readers preferred ILUMENATE images in 22/22 cases and scored ILUMENATE superiorly for overall quality, sharpness, and noise ( p < 0.05). Conclusions ILUMENATE enhanced image resolution, resulting in improved overall image quality, reduced calcium blooming artifacts, and improved lumen visibility in cCTA exams performed using EID‐CT. This could potentially allow for improved accessibility to UHR image quality, allowing for more accurate assessment of CAD.

Abstract Background Accurate low‐dose CT simulation is required to efficiently assess reconstruction and dose reduction techniques. Projection domain noise insertion requires proprietary information from manufacturers. Analytic image domain noise insertion methods are successful for linear reconstruction algorithms, however extending them to non‐linear algorithms remains challenging. Emerging, deep‐learning‐based image domain noise insertion methods have potential, but few approaches have explicitly incorporated physics information and a texture‐synthesis model to guide the generation of locally and globally correlated noise texture. Purpose We proposed a p hysics‐informed model‐based gener a tive neura l network for simulating scann e r‐ and algori t hm‐specific low‐dose C T e xams (PALETTE). It is expected to provide an alternative to projection domain noise insertion methods in the absence of manufacturers’ proprietary information and tools. Methods PALETTE integrated a physics‐based noise prior generation process, a Noise2Noisier sub‐network, and a noise texture synthesis sub‐network. The Noise2Noisier sub‐network provided a bias prior, which, combined with the noise prior, served as the inputs to noise texture synthesis sub‐network. Explicit regularizations in spatial and frequency domains were developed to account for noise spatial correlation and frequency characteristics. For proof‐of‐concept, PALETTE was trained and validated for a commercial iterative reconstruction algorithm (SAFIRE, Siemens Healthineers), using the paired routine and 25% dose images from CT phantoms (lateral size 30–40 cm; three training and four testing phantoms) and open‐access patient cases (10 training and 20 testing cases). In phantom validation, noise power spectra (NPS) were compared in water background and tissue‐mimicking inserts, using peak frequency and mean‐absolute‐error (MAE). In patient case evaluation, visual inspection and quantitative assessment were conducted on axial, coronal, and sagittal planes. Local and global noise texture were visually inspected in low‐dose CT images and the difference images between routine and low dose. Noise levels in liver and fat were measured. Local and global 2D Fourier magnitude spectra of the difference images and the corresponding radial mean profiles were used to assess similarity in noise frequency components within tissues and entire field‐of‐view, using spectral correlation mapper (SCM) and spectral angle mapper (SAM). Several baseline neural network models (e.g., GAN) were included in the evaluation. Statistical significance was tested using a t ‐test for related samples. Results PALETTE‐derived NPS showed accurate noise peak frequency (PALETTE/reference: water 1.40/1.40 lp/cm; inserts 1.7/1.7lp/cm) and small MAE (≤0.65 HU 2 cm 2 ). PALETTE created anatomy‐dependent noise texture, showing realistic local and global granularity and streaks. No statistically significant difference was observed in noise levels ( p > 0.05). Noise range was comparable across 3D image volume (PALETTE / reference):liver – [18.0, 53.4] / [19.3, 50.0] HU; fat – [11.7, 42.4] / [12.1, 41.3] HU. Percent absolute difference of local noise was small (mean ± standard deviation): liver 4.1%±3.1%, fat 4.6%±3.1%. Noise frequency distribution was close to the reference (mean per case): SCM ≥ 0.92, SAM ≤ 0.22. Additionally, PALETTE outperformed all baseline models in visual inspection and quantitative comparison. Conclusion PALETTE can provide high‐quality image domain noise insertion for simulating accurate low‐dose CT images created with a commercial non‐linear reconstruction algorithm.

AbstractBackgroundDeep‐learning‐based image reconstruction and noise reduction methods (DLIR) have been increasingly deployed in clinical CT. Accurate image quality assessment of these methods is challenging as the performance measured using physical phantoms may not represent the true performance of DLIR in patients since DLIR is trained mostly on patient images.PurposeIn this work, we aim to develop a patient‐data‐based virtual imaging trial framework and, as a first application, use it to measure the spatial resolution properties of a DLIR method.MethodsThe patient‐data‐based virtual imaging trial framework consists of five steps: (1) insertion of lesions into projection domain data using the acquisition geometry of the patient exam to simulate different lesion characteristics; (2) insertion of noise into projection domain data using a realistic photon statistical model of the CT system to simulate different dose levels; (3) creation of DLIR‐processed images from projection or image data; (4) creation of ensembles of DLIR‐processed patient images from a large number of noise and lesion realizations; and (5) evaluation of image quality using ensemble DLIR images. This framework was applied to measure the spatial resolution of a ResNet based deep convolutional neural network (DCNN) trained on patient images. Lesions in a cylindrical shape and different contrast levels (−500, −100, −50, −20, −10 HU) were inserted to the lower right lobe of the liver in a patient case. Multiple dose levels were simulated (50%, 25%, 12.5%). Each lesion and dose condition had 600 noise realizations. Multiple reconstruction and denoising methods were used on all the noise realizations, including the original filtered‐backprojection (FBP), iterative reconstruction (IR), and the DCNN method with three different strength setting (DCNN‐weak, DCNN‐medium, and DCNN‐strong). Mean lesion signal was calculated by performing ensemble averaging of all the noise realizations for each lesion and dose condition and then subtracting the lesion‐present images from the lesion absent images. Modulation transfer functions (MTFs) both in‐plane and along the z‐axis were calculated based on the mean lesion signals. The standard deviations of MTFs at each condition were estimated with bootstrapping: randomly sampling (with replacement) all the DLIR/FBP/IR images from the ensemble data (600 samples) at each condition. The impact of varying lesion contrast, dose levels, and denoising strengths were evaluated. Statistical analysis with paired t‐test was used to compare the z‐axis and in‐plane spatial resolution of five algorithms for five different contrasts and three dose levels.ResultsThe in‐plane and z‐axis spatial resolution degradation of DCNN becomes more severe as the contrast or radiation dose decreased, or DCNN denoising strength increased. In comparison with FBP, a 59.5% and 4.1% reduction of in‐plane and z‐axis MTF (in terms of spatial frequencies at 50% MTF), respectively, was observed at low contrast (−10 HU) for DCNN with the highest denoising strength at 25% routine dose level. When the dose level reduces from 50% to 12.5% of routine dose, the in‐plane and z‐axis MTFs reduces from 92.1% to 76.3%, and from 98.9% to 95.5%, respectively, at contrast of −100 HU, using FBP as the reference. For most conditions of contrasts and dose levels, significant differences were found among the five algorithms, with the following relationship in both in‐plane and cross‐plane spatial resolution: FBP > DCNN‐Weak > IR > DCNN‐Medium > DCNN‐Strong. The spatial resolution difference among algorithms decreases at higher contrast or dose levels.ConclusionsA patient‐data‐based virtual imaging trial framework was developed and applied to measuring the spatial resolution properties of a DCNN noise reduction method at different contrast and dose levels using real patient data. As with other non‐linear image reconstruction and post‐processing techniques, the evaluated DCNN method degraded the in‐plane and z‐axis spatial resolution at lower contrast levels, lower radiation dose, and higher denoising strength.

Abstract Objectives. To improve quality of coronary CT angiography (CCTA) images using a generalizable motion-correction algorithm. Approach . A neural network with attention gate and spatial transformer (ATOM) was developed to correct coronary motion. Phantom and patient CCTA images (39 males, 32 females, age range 19–92, scan date 02/2020 to 10/2021) retrospectively collected from dual-source CT were used to create training, development, and testing sets corresponding to 140- and 75 ms temporal resolution, with 75 ms images as labels. To test generalizability, ATOM was deployed for locally adaptive motion-correction in both 140- and 75 ms patient images. Objective metrics were used to assess motion-corrupted and corrected phantom and patient images, including structural-similarity-index (SSIM), dice-similarity-coefficient (DSC), peak-signal-noise-ratio (PSNR), and normalized root-mean-square-error (NRMSE). In objective quality assessment, ATOM was compared with several baseline networks, including U-net, U-net plus attention gate, U-net plus spatial transformer, VDSR, and ResNet. Two cardiac radiologists independently interpreted motion-corrupted and -corrected images at 75 and 140 ms in a blinded fashion and ranked diagnostic image quality (worst to best: 1–4, no ties). Main results . ATOM improved quality metrics ( p < 0.05) before/after correction: in phantom, SSIM 0.87/0.95, DSC 0.85/0.93, PSNR 19.4/22.5, NRMSE 0.38/0.27; in patient images, SSIM 0.82/0.88, DSC 0.88/0.90, PSNR 30.0/32.0, NRMSE 0.16/0.12. ATOM provided more consistent improvement of objective image quality, compared to the presented baseline networks. The motion-corrected images received better ranks than un-corrected at the same temporal resolution ( p < 0.05): 140 ms images 1.65/2.25, and 75 ms images 3.1/3.2. The motion-corrected 75 ms images received the best rank in 65% of testing cases. A fair-to-good inter-reader agreement was observed (Kappa score 0.58). Significance . ATOM reduces motion artifacts, improving visualization of coronary arteries. This algorithm can be used to virtually improve temporal resolution in both single- and dual-source CT.

AbstractBackgroundPhoton‐counting‐detector CT (PCD‐CT) enables the production of virtual monoenergetic images (VMIs) at a high spatial resolution (HR) via simultaneous acquisition of multi‐energy data. However, noise levels in these HR VMIs are markedly increased.PurposeTo develop a deep learning technique that utilizes a lower noise VMI as prior information to reduce image noise in HR, PCD‐CT coronary CT angiography (CTA).MethodsCoronary CTA exams of 10 patients were acquired using PCD‐CT (NAEOTOM Alpha, Siemens Healthineers). A prior‐information‐enabled neural network (Pie‐Net) was developed, treating one lower‐noise VMI (e.g., 70 keV) as a prior input and one noisy VMI (e.g., 50 keV or 100 keV) as another. For data preprocessing, noisy VMIs were reconstructed by filtered back‐projection (FBP) and iterative reconstruction (IR), which were then subtracted to generate “noise‐only” images. Spatial decoupling was applied to the noise‐only images to mitigate overfitting and improve randomization. Thicker slice averaging was used for the IR and prior images. The final training inputs for the convolutional neural network (CNN) inside the Pie‐Net consisted of thicker‐slice signal images with the reinsertion of spatially decoupled noise‐only images and the thicker‐slice prior images. The CNN training labels consisted of the corresponding thicker‐slice label images without noise insertion. Pie‐Net's performance was evaluated in terms of image noise, spatial detail preservation, and quantitative accuracy, and compared to a U‐net‐based method that did not include prior information.ResultsPie‐Net provided strong noise reduction, by 95 ± 1% relative to FBP and by 60 ± 8% relative to IR. For HR VMIs at different keV (e.g., 50 keV or 100 keV), Pie‐Net maintained spatial and spectral fidelity. The inclusion of prior information from the PCD‐CT data in the spectral domain was able to improve a robust deep learning‐based denoising performance compared to the U‐net‐based method, which caused some loss of spatial detail and introduced some artifacts.ConclusionThe proposed Pie‐Net achieved substantial noise reduction while preserving HR VMI's spatial and spectral properties.