Leng, Shuai

PurposeFiltering measured projections with a particular convolutional kernel is an essential step in analytic reconstruction of computed tomography (CT) images. A tradeoff between noise and spatial resolution exists for different choices of reconstruction kernel. In a clinical setting, this often requires producing multiple images reconstructed with different kernels for a single CT exam, which increases the burden of computation, networking, archival, and reading. We address this problem by training a deep convolutional neural network (CNN) to synthesize multiple input images into a single output image which exhibits low noise while also preserving features in images reconstructed with the sharpest kernels.MethodsA CNN architecture consisting of repeated blocks of residual units containing a total of 20 convolutional layers was used to combine features. The CNN inputs consisted of two images produced with different reconstruction kernels, one smooth and one sharp, which were stacked in the channel dimension. The network was trained using supervised learning with both full‐dose and simulated quarter‐dose abdominal CT images. After training, the performance was evaluated using a reserved set of full‐dose scans that were not used for network optimization. Noise reduction performance was measured by comparing root mean square (RMS) measurements in uniform regions. Spatial resolution was compared using line profiles of anatomic features.ResultsFor the regions tested, the synthetic images feature noise levels slightly below those of the smooth input images, while maintaining the resolution of anatomic details found in the sharp input images.ConclusionsA deep CNN can be used combine features from CT images reconstructed with different kernels to produce a single synthesized image series that exhibits both low noise and high spatial resolution. This approach has implications for improving image quality, reducing radiation dose, and simplifying the clinical workflow for CT imaging.

Background The rapid development and complexity of new x‐ray computed tomography (CT) technologies and the need for evidence‐based optimization of image quality with respect to radiation and contrast media dose call for an updated approach towards CT performance evaluation. Aims This report offers updated testing guidelines for testing CT systems with an enhanced focus on the operational performance including iterative reconstructions and automatic exposure control (AEC) techniques. Materials and Methods The report was developed based on a comprehensive review of best methods and practices in the scientific literature. The detailed methods include the assessment of 1) CT noise (magnitude, texture, nonuniformity, inhomogeneity), 2) resolution (task transfer function under varying conditions and its scalar reflections), 3) task‐based performance (detectability, estimability), and 4) AEC performance (spatial, noise, and mA concordance of attenuation and exposure modulation). The methods include varying reconstruction and tube current modulation conditions, standardized testing protocols, and standardized quantities and metrology to facilitate tracking, benchmarking, and quantitative comparisons. Results The methods, implemented in cited publications, are robust to provide a representative reflection of CT system performance as used operationally in a clinical facility. The methods include recommendations for phantoms and phantom image analysis. Discussion In line with the current professional trajectory of the field toward quantitation and operational engagement, the stated methods offer quantitation that is more predictive of clinical performance than specification‐based approaches. They can pave the way to approach performance testing of new CT systems not only in terms of acceptance testing (i.e., verifying a device meets predefined specifications), but also system commissioning (i.e., determining how the system can be used most effectively in clinical practice). Conclusion We offer a set of common testing procedures that can be utilized towards the optimal clinical utilization of CT imaging devices, benchmarking across varying systems and times, and a basis to develop future performance‐based criteria for CT imaging.

PurposePhoton‐counting‐detector‐computed tomography (PCD‐CT) allows separation of multiple, simultaneously imaged contrast agents, such as iodine (I), gadolinium (Gd), and bismuth (Bi). However, PCDs suffer from several technical limitations such as charge sharing, K‐edge escape, and pulse pile‐up, which compromise spectral separation of multi‐energy data and degrade multi‐contrast imaging performance. The purpose of this work was to determine the performance of a dual‐source (DS) PCD‐CT relative to a single‐source (SS) PCD‐CT for the separation of simultaneously imaged I, Gd, and Bi contrast agents.MethodsPhantom experiments were performed using a research whole‐body PCD‐CT and head/abdomen‐sized phantoms containing vials of different I, Gd, Bi concentrations. To emulate a DS‐PCD‐CT, the phantoms were scanned twice on the SS‐PCD‐CT using different tube potentials for each scan. A tube potential of 80 kV (energy thresholds = 25/50 keV) was used for low‐energy tube, while the high‐energy tube used Sn140 kV (Sn indicates tin filter) and thresholds of 25/90 keV. The same phantoms were scanned also on the SS‐PCD‐CT using the chess acquisition mode. In chess mode, the 4 × 4 subpixels within a macro detector pixel are split into two sets based on a chess‐board pattern. With each subpixel set having two energy thresholds, chess mode allows four energy‐bin data sets, which permits simultaneous multi‐contrast imaging. Because of this design, only 50% area of each detector pixel is configured to receive photons of a pre‐defined threshold, leading to 50% dose utilization efficiency. To compensate for this dose inefficiency, the radiation dose for this scan was doubled compared to DS‐PCD‐CT. A 140 kV tube potential and thresholds = 25/50/75/90 keV were used. These settings were determined based on the K‐edges of Gd, and Bi, and were found to yield good differentiation of I/Gd/Bi based on phantom experiments and other literature. The energy‐bin images obtained from each scan (scan pair) were used to generate I‐, Gd‐, Bi‐specific image via material decomposition. Root‐mean‐square‐error (RMSE) between the known and measured concentrations was calculated for each scenario. A 20‐cm water cylinder phantom was scanned on both systems, which was used for evaluating the magnitude of noise, and noise power spectra (NPS) of I/Gd/Bi‐specific images.ResultsPhantom results showed that DS‐PCD‐CT reduced noise in material‐specific images for both head and body phantoms compared to SS‐PCD‐CT. The noise level of SS‐PCD was reduced from 2.55 to 0.90 mg/mL (I), 1.97 to 0.78 mg/mL (Gd), and 0.85 to 0.74 mg/mL (Bi) using DS‐PCD. NPS analysis showed that the noise texture of images acquired on both systems is similar. For the body phantom, the RMSE for SS‐PCD‐CT was reduced relative to DS‐PCD‐CT from 10.52 to 2.76 mg/mL (I), 7.90 to 2.01 mg/mL (Gd), and 1.91 to 1.16 mg/mL (Bi). A similar trend was observed for the head phantom: RMSE reduced from 2.59 (SS‐PCD) to 0.72 (DS‐PCD) mg/mL (I), 2.02 to 0.58 mg/mL (Gd), and 0.85 to 0.57 mg/mL (Bi).ConclusionWe demonstrate the feasibility of performing simultaneous imaging of I, Gd, and Bi materials on DS‐PCD‐CT. Under the condition without cross scattering, DS‐PCD reduced the RMSE for quantification of material concentration in relative to a SS‐PCD‐CT system using chess mode.

PurposeThis work aims to develop a new framework of image quality assessment using deep learning‐based model observer (DL‐MO) and to validate it in a low‐contrast lesion detection task that involves CT images with patient anatomical background.MethodsThe DL‐MO was developed using the transfer learning strategy to incorporate a pretrained deep convolutional neural network (CNN), a partial least square regression discriminant analysis (PLS‐DA) model and an internal noise component. The CNN was previously trained to achieve the state‐of‐the‐art classification accuracy over a natural image database. The earlier layers of the CNN were used as a deep feature extractor, with the assumption that similarity exists between the CNN and the human visual system. The PLSR model was used to further engineer the deep feature for the lesion detection task in CT images. The internal noise component was incorporated to model the inefficiency and variability of human observer (HO) performance, and to generate the ultimate DL‐MO test statistics. Seven abdominal CT exams were retrospectively collected from the same type of CT scanners. To compare DL‐MO with HO, 12 experimental conditions with varying lesion size, lesion contrast, radiation dose, and reconstruction types were generated, each condition with 154 trials. CT images of a real liver metastatic lesion were numerically modified to generate lesion models with four lesion sizes (5, 7, 9, and 11 mm) and three contrast levels (15, 20, and 25 HU). The lesions were inserted into patient liver images using a projection‐based method. A validated noise insertion tool was used to synthesize CT exams with 50% and 25% of routine radiation dose level. CT images were reconstructed using the weighted filtered back projection algorithm and an iterative reconstruction algorithm. Four medical physicists performed a two‐alternative forced choice (2AFC) detection task (with multislice scrolling viewing mode) on patient images across the 12 experimental conditions. DL‐MO was operated on the same datasets. Statistical analyses were performed to evaluate the correlation and agreement between DL‐MO and HO.ResultsA statistically significant positive correlation was observed between DL‐MO and HO for the 2AFC low‐contrast detection task that involves patient liver background. The corresponding Pearson product moment correlation coefficient was 0.986 [95% confidence interval (0.950, 0.996)]. Bland–Altman agreement analysis did not indicate statistically significant differences.ConclusionsThe proposed DL‐MO is highly correlated with HO in a low‐contrast detection task that involves realistic patient liver background. This study demonstrated the potential of the proposed DL‐MO to assess image quality directly based on patient images in realistic, clinically relevant CT tasks.

Objectives The aim of this study was to investigate the impact on metal artifacts and dose efficiency of using a tin filter in combination with high-energy threshold (TH) images of a photon-counting detector (PCD) computed tomography (CT) system. Materials and Methods A 3D-printed spine with pedicle screws was scanned on a PCD-CT system with and without tin filtration. Image noise and severity of artifacts were measured for low-energy threshold (TL) and TH images. In a prospective, institutional review board–approved, Health Insurance Portability and Accountability Act-compliant study, 20 patients having a clinical energy-integrating detector (EID) CT were scanned on a PCD-CT system using tin filtration. Images were reviewed by 3 radiologists to evaluate visualization of anatomic structures, diagnostic confidence, and image preference. Artifact severity and image noise were measured. Wilcoxon signed rank was used to test differences between PCD-CT TH and EID-CT images. Results Phantom TH images with tin filtration reduced metal artifacts and had comparable noise (32 HU) to TL images (29 HU) acquired without tin filtration. Visualization scores for the cortex, trabeculae, and implant-trabecular interface from PCD-CT TH images (4.4 ± 0.9, 4.4 ± 1.0, and 4.4 ± 1.0) were significantly higher (P < 0.0001) than EID-CT images (3.3 ± 1.3, 3.3 ± 1.2, and 3.3 ± 1.6). A strong preference was shown for PCD-CT TH images due to improved diagnostic confidence and decreased artifact severity. Noise in PCD-CT TH images (93 ± 41 HU) was significantly lower than that in EID-CT images (133 ± 92 HU, P < 0.05). Conclusions Threshold high images acquired with tin filtration on PCD-CT demonstrated a substantial decrease in metal artifacts and an increase in dose efficiency compared with EID-CT.

Objectives The aim of this study was to evaluate if a high-resolution photon-counting detector computed tomography (PCD-CT) system with a 1024×1024 matrix reconstruction can improve the visualization of fine structures in the lungs compared with conventional high-resolution CT (HRCT). Materials and Methods Twenty-two adult patients referred for clinical chest HRCT (mean CTDI vol, 13.58 mGy) underwent additional dose-matched PCD-CT (mean volume CT dose index, 13.37 mGy) after written informed consent. Computed tomography images were reconstructed at a slice thickness of 1.5 mm and an image increment of 1 mm with our routine HRCT reconstruction kernels (B46 and Bv49) at 512 and 1024 matrix sizes for conventional energy-integrating detector (EID) CT scans. For PCD-CT, routine B46 kernel and an additional sharp kernel (Q65, unavailable for EID) images were reconstructed at 1024 matrix size. Two thoracic radiologists compared images from EID and PCD-CT noting the highest level bronchus clearly identified in each lobe of the right lung, and rating bronchial wall conspicuity of third- and fourth-order bronchi. Lung nodules were also compared with the B46/EID/512 images using a 5-point Likert scale. Statistical analysis was performed using a Wilcoxon signed rank test with a P < 0.05 considered significant. Results Compared with B46/EID/512, readers detected higher-order bronchi using B46/PCD/1024 and Q65/PCD/1024 images for every lung lobe (P < 0.0015), but in only the right middle lobe for B46/EID/1024 (P = 0.007). Readers were able to better identify bronchial walls of the third- and fourth-order bronchi better using the Q65/PCD/1024 images (mean Likert scores of 1.1 and 1.5), which was significantly higher compared with B46/EID/1024 or B46/PCD/1024 images (mean difference, 0.8; P < 0.0001). The Q65/PCD/1024 images had a mean nodule score of 1 ± 1.3 for reader 1, and −0.1 (0.9) for reader 2, with one reader having improved nodule evaluation scores for both PCD kernels (P < 0.001), and the other reader not identifying any increased advantage over B46/EID/1024 (P = 1.0). Conclusions High-resolution lung PCD-CT with 1024 image matrix reconstruction increased radiologists' ability to visualize higher-order bronchi and bronchial walls without compromising nodule evaluation compared with current chest CT, creating an opportunity for radiologists to better evaluate airway pathology.