Gong, Hao

AbstractBackgroundDeep artificial neural networks such as convolutional neural networks (CNNs) have been shown to be effective models for reducing noise in CT images while preserving anatomic details. A practical bottleneck for developing CNN‐based denoising models is the procurement of training data consisting of paired examples of high‐noise and low‐noise CT images. Obtaining these paired data are not practical in a clinical setting where the raw projection data is not available. This work outlines a technique to optimize CNN denoising models using methods that are available in a routine clinical setting.PurposeTo demonstrate a phantom‐based training framework for CNN noise reduction that can be efficiently implemented on any CT scanner.MethodsThe phantom‐based training framework uses supervised learning in which training data are synthesized using an image‐based noise insertion technique. Ten patient image series were used for training and validation (9:1) and noise‐only images obtained from anthropomorphic phantom scans. Phantom noise‐only images were superimposed on patient images to imitate low‐dose CT images for use in training. A modified U‐Net architecture was used with mean‐squared‐error and feature reconstruction loss. The training framework was tested for clinically indicated whole‐body‐low‐dose CT images, as well as routine abdomen‐pelvis exams for which projection data was unavailable. Performance was assessed based on root‐mean‐square error, structural similarity, line profiles, and visual assessment.ResultsWhen the CNN was tested on five reserved quarter‐dose whole‐body‐low‐dose CT images, noise was reduced by 75%, root‐mean‐square‐error reduced by 34%, and structural similarity increased by 60%. Visual analysis and line profiles indicated that the method significantly reduced noise while maintaining spatial resolution of anatomic features.ConclusionThe proposed phantom‐based training framework demonstrated strong noise reduction while preserving spatial detail. Because this method is based within the image domain, it can be easily implemented without access to projection data.

The scaling of respiratory structures has been hypothesized to be a major driving factor in the evolution of many aspects of animal physiology. Here, we provide the first assessment of the scaling of the spiracles in insects using 10 scarab beetle species differing 180× in mass, including some of the most massive extant insect species. Using X-ray microtomography, we measured the cross-sectional area and depth of all eight spiracles, enabling the calculation of their diffusive and advective capacities. Each of these metrics scaled with geometric isometry. Because diffusive capacities scale with lower slopes than metabolic rates, the largest beetles measured require 10-fold higher P O2 gradients across the spiracles to sustain metabolism by diffusion compared to the smallest species. Large beetles can exchange sufficient oxygen for resting metabolism by diffusion across the spiracles, but not during flight. In contrast, spiracular advective capacities scale similarly or more steeply than metabolic rates, so spiracular advective capacities should match or exceed respiratory demands in the largest beetles. These data illustrate a general principle of gas exchange: scaling of respiratory transport structures with geometric isometry diminishes the potential for diffusive gas exchange but enhances advective capacities; combining such structural scaling with muscle-driven ventilation allows larger animals to achieve high metabolic rates when active.

AbstractBackgroundDual‐energy CT with virtual noncalcium (VNCa) images allows the evaluation of focal intramedullary bone marrow involvement in patients with multiple myeloma. However, current commercial VNCa techniques suffer from excessive image noise and artifacts due to material decomposition used in synthesizing VNCa images.ObjectivesIn this work, we aim to improve VNCa image quality for the assessment of focal multiple myeloma, using an Artificial intelligence based Generalizable Algorithm for mulTi‐Energy CT (AGATE) method.Materials and methodsAGATE method used a custom dual‐task convolutional neural network (CNN) that concurrently carries out material classification and quantification. The material classification task provided an auxiliary regularization to the material quantification task. CNN parameters were optimized using custom loss functions that involved cross‐entropy, physics‐informed constraints, structural redundancy in spectral and material images, and texture information in spectral images. For training data, CT phantoms (diameters 30 to 45 cm) with tissue‐mimicking inserts were scanned on a third generation dual‐source CT system. Scans were performed at routine dose and half of the routine dose. Small image patches (i.e., 40 × 40 pixels) of tissue‐mimicking inserts with known basis material densities were extracted for training samples. Numerically simulated insert materials with various shapes increased diversity of training samples. Generalizability of AGATE was evaluated using CT images from phantoms and patients. In phantoms, material decomposition accuracy was estimated using mean‐absolute‐percent‐error (MAPE), using physical inserts that were not used during the training. Noise power spectrum (NPS) and modulation transfer function (MTF) were compared across phantom sizes and radiation dose levels. Five patients with multiple myeloma underwent dual‐energy CT, with VNCa images generated using a commercial method and AGATE. Two fellowship‐trained musculoskeletal radiologists reviewed the VNCa images (commercial and AGATE) side‐by‐side using a dual‐monitor display, blinded to VNCa type, rating the image quality for focal multiple myeloma lesion visualization using a 5‐level Likert comparison scale (−2 = worse visualization and diagnostic confidence, −1 = worse visualization but equivalent diagnostic confidence, 0 = equivalent visualization and diagnostic confidence, 1 = improved visualization but equivalent diagnostic confidence, 2 = improved visualization and diagnostic confidence). A post hoc assignment of comparison ratings was performed to rank AGATE images in comparison to commercial ones.ResultsAGATE demonstrated consistent material quantification accuracy across phantom sizes and radiation dose levels, with MAPE ranging from 0.7% to 4.4% across all testing materials. Compared to commercial VNCa images, the AGATE‐synthesized VNCa images yielded considerably lower image noise (50–77% noise reduction) without compromising noise texture or spatial resolution across different phantom sizes and two radiation doses. AGATE VNCa images had markedly reduced area under NPS curves and maintained NPS peak frequency (0.7 lp/cm to 1.0 lp/cm), with similar MTF curves (50% MTF at 3.0 lp/cm). In patients, AGATE demonstrated reduced image noise and artifacts with improved delineation of focal multiple myeloma lesions (all readers comparison scores indicating improved overall diagnostic image quality [scores 1 or 2]).ConclusionsAGATE demonstrated reduced noise and artifacts in VNCa images and ability to improve visualization of bone marrow lesions for assessing multiple myeloma.

AbstractPurposeConventional model observers (MO) in CT are often limited to a uniform background or varying background that is random and can be modeled in an analytical form. It is unclear if these conventional MOs can be readily generalized to predict human observer performance in clinical CT tasks that involve realistic anatomical background. Deep‐learning‐based model observers (DL‐MO) have recently been developed, but have not been validated for challenging low contrast diagnostic tasks in abdominal CT. We consequently sought to validate a DL‐MO for a low‐contrast hepatic metastases localization task.MethodsWe adapted our recently developed DL‐MO framework for the liver metastases localization task. Our previously‐validated projection‐domain lesion‐/noise‐insertion techniques were used to synthesize realistic positive and low‐dose abdominal CT exams, using the archived patient projection data. Ten experimental conditions were generated, which involved different lesion sizes/contrasts, radiation dose levels, and image reconstruction types. Each condition included 100 trials generated from a patient cohort of 7 cases. Each trial was presented as liver image patches (160×160×5 voxels). The DL‐MO performance was calculated for each condition and was compared with human observer performance, which was obtained by three sub‐specialized radiologists in an observer study. The performance of DL‐MO and radiologists was gauged by the area under localization receiver‐operating‐characteristic curves. The generalization performance of the DL‐MO was estimated with the repeated twofold cross‐validation method over the same set of trials used in the human observer study. A multi‐slice Channelized Hoteling Observers (CHO) was compared with the DL‐MO across the same experimental conditions.ResultsThe performance of DL‐MO was highly correlated to that of radiologists (Pearson's correlation coefficient: 0.987; 95% CI: [0.942, 0.997]). The performance level of DL‐MO was comparable to that of the grouped radiologists, that is, the mean performance difference was ‐3.3%. The CHO performance was poorer than the grouped radiologist performance, before internal noise could be added. The correlation between CHO and radiologists was weaker (Pearson's correlation coefficient: 0.812, and 95% CI: [0.378, 0.955]), and the corresponding performance bias (‐29.5%) was statistically significant.ConclusionThe presented study demonstrated the potential of using the DL‐MO for image quality assessment in patient abdominal CT tasks.

Objectives: The objectives of the study were to estimate the impact of high matrix image reconstruction on chest computed tomography (CT) compared to standard image reconstruction. Material and Methods: This retrospective study included patients with interstitial or parenchymal lung disease, airway disease, and pulmonary nodules who underwent chest CT. Chest CT images were reconstructed using high matrix (1024 × 1024) or standard matrix (512 × 512), with all other parameters matched. Two radiologists, blinded to reconstruction technique, independently examined each lung, viewing image sets side by side and rating the conspicuity of imaging findings using a 5-point relative conspicuity scale. The presence of pulmonary nodules and confidence in classification of internal attenuation was also graded. Overall image quality and subjective noise/artifacts were assessed. Results: Thirty-four patients with 68 lungs were evaluated. Relative conspicuity scores were significantly higher using high matrix image reconstruction for all imaging findings indicative of idiopathic lung fibrosis (peripheral airway visualization, interlobular septal thickening, intralobular reticular opacity, and end-stage fibrotic change; P ≤ 0.001) along with emphysema, mosaic attenuation, and fourth order bronchi for both readers (P ≤ 0.001). High matrix reconstruction did not improve confidence in the presence or classification of internal nodule attenuation for either reader. Overall image quality was increased but not subjective noise/artifacts with high matrix image reconstruction for both readers (P < 0.001). Conclusion: High matrix image reconstruction significantly improves the conspicuity of imaging findings reflecting interstitial lung disease and may be useful for diagnosis or treatment response assessment.

AbstractPurposeEye‐tracking approaches have been used to understand the visual search process in radiology. However, previous eye‐tracking work in computer tomography (CT) has been limited largely to single cross‐sectional images or video playback of the reconstructed volume, which do not accurately reflect radiologists' visual search activities and their interactivity with three‐dimensional image data at a computer workstation (e.g., scroll, pan, and zoom) for visual evaluation of diagnostic imaging targets. We have developed a platform that integrates eye‐tracking hardware with in‐house‐developed reader workstation software to allow monitoring of the visual search process and reader‐image interactions in clinically relevant reader tasks. The purpose of this work is to validate the spatial accuracy of eye‐tracking data using this platform for different eye‐tracking data acquisition modes.MethodsAn eye‐tracker was integrated with a previously developed workstation designed for reader performance studies. The integrated system captured real‐time eye movement and workstation events at 1000 Hz sampling frequency. The eye‐tracker was operated either in head‐stabilized mode or in free‐movement mode. In head‐stabilized mode, the reader positioned their head on a manufacturer‐provided chinrest. In free‐movement mode, a biofeedback tool emitted an audio cue when the head position was outside the data collection range (general biofeedback) or outside a narrower range of positions near the calibration position (strict biofeedback). Four radiologists and one resident were invited to participate in three studies to determine eye‐tracking spatial accuracy under three constraint conditions: head‐stabilized mode (i.e., with use of a chin rest), free movement with general biofeedback, and free movement with strict biofeedback. Study 1 evaluated the impact of head stabilization versus general or strict biofeedback using a cross‐hair target prior to the integration of the eye‐tracker with the image viewing workstation. In Study 2, after integration of the eye‐tracker and reader workstation, readers were asked to fixate on targets that were randomly distributed within a volumetric digital phantom. In Study 3, readers used the integrated system to scroll through volumetric patient CT angiographic images while fixating on the centerline of designated blood vessels (from the left coronary artery to dorsalis pedis artery). Spatial accuracy was quantified as the offset between the center of the intended target and the detected fixation using units of image pixels and the degree of visual angle.ResultsThe three head position constraint conditions yielded comparable accuracy in the studies using digital phantoms. For Study 1 involving the digital crosshairs, the median ± the standard deviation of offset values among readers were 15.2 ± 7.0 image pixels with the chinrest, 14.2 ± 3.6 image pixels with strict biofeedback, and 19.1 ± 6.5 image pixels with general biofeedback. For Study 2 using the random dot phantom, the median ± standard deviation offset values were 16.7 ± 28.8 pixels with use of a chinrest, 16.5 ± 24.6 pixels using strict biofeedback, and 18.0 ± 22.4 pixels using general biofeedback, which translated to a visual angle of about 0.8° for all three conditions. We found no obvious association between eye‐tracking accuracy and target size or view time. In Study 3 viewing patient images, use of the chinrest and strict biofeedback demonstrated comparable accuracy, while the use of general biofeedback demonstrated a slightly worse accuracy. The median ± standard deviation of offset values were 14.8 ± 11.4 pixels with use of a chinrest, 21.0 ± 16.2 pixels using strict biofeedback, and 29.7 ± 20.9 image pixels using general biofeedback. These corresponded to visual angles ranging from 0.7° to 1.3°.ConclusionsAn integrated eye‐tracker system to assess reader eye movement and interactive viewing in relation to imaging targets demonstrated reasonable spatial accuracy for assessment of visual fixation. The head‐free movement condition with audio biofeedback performed similarly to head‐stabilized mode.

AbstractObjectiveIn X‐raycomputed tomography (CT), many important clinical applications may benefit from a fast acquisition speed. The helical scan is the most widely used acquisition mode in clinical CT, where a fast helical pitch can improve the acquisition speed. However, on a typical single‐source helical CT (SSCT) system, the helical pitch p typically cannot exceed 1.5; otherwise, reconstruction artifacts will result from data insufficiency. The purpose of this work is to develop a deep convolutional neural network (CNN) to correct for artifacts caused by an ultra‐fast pitch, which can enable faster acquisition speed than what is currently achievable.MethodsA customized CNN (denoted as ultra‐fast‐pitch network (UFP‐net)) was developed to restore the underlying anatomical structure from the artifact‐corrupted post‐reconstruction data acquired from SSCT with ultra‐fast pitch (i.e., p ≥ 2). UFP‐net employed residual learning to capture the features of image artifacts. UFP‐net further deployed in‐house‐customized functional blocks with spatial‐domain local operators and frequency‐domain non‐local operators, to explore multi‐scale feature representation. Images of contrast‐enhanced patient exams (n = 83) with routine pitch setting (i.e., p < 1) were retrospectively collected, which were used as training and testing datasets. This patient cohort involved CT exams over different scan ranges of anatomy (chest, abdomen, and pelvis) and CT systems (Siemens Definition, Definition Flash, Definition AS+, Siemens Healthcare, Inc.), and the corresponding base CT scanning protocols used consistent settings of major scan parameters (e.g., collimation and pitch). Forward projection of the original images was calculated to synthesize helical CT scans with one regular pitch setting (p = 1) and two ultra‐fast‐pitch setting (p = 2 and 3). All patient images were reconstructed using the standard filtered‐back‐projection (FBP) algorithm. A customized multi‐stage training scheme was developed to incrementally optimize the parameters of UFP‐net, using ultra‐fast‐pitch images as network inputs and regular pitch images as labels. Visual inspection was conducted to evaluate image quality. Structural similarity index (SSIM) and relative root‐mean‐square error (rRMSE) were used as quantitative quality metrics.ResultsThe UFP‐net dramatically improved image quality over standard FBP at both ultra‐fast‐pitch settings. At p = 2, UFP‐net yielded higher mean SSIM (> 0.98) with lower mean rRMSE (< 2.9%), compared to FBP (mean SSIM < 0.93; mean rRMSE > 9.1%). Quantitative metrics at p = 3: UFP‐net—mean SSIM [0.86, 0.94] and mean rRMSE [5.0%, 8.2%]; FBP—mean SSIM [0.36, 0.61] and mean rRMSE [36.0%, 58.6%].ConclusionThe proposed UFP‐net has the potential to enable ultra‐fast data acquisition in clinical CT without sacrificing image quality. This method has demonstrated reasonable generalizability over different body parts when the corresponding CT exams involved consistent base scan parameters.

PurposeTo develop a convolutional neural network (CNN) that can directly estimate material density distribution from multi‐energy computed tomography (CT) images without performing conventional material decomposition.MethodsThe proposed CNN (denoted as Incept‐net) followed the general framework of encoder–decoder network, with an assumption that local image information was sufficient for modeling the nonlinear physical process of multi‐energy CT. Incept‐net was implemented with a customized loss function, including an in‐house‐designed image‐gradient‐correlation (IGC) regularizer to improve edge preservation. The network consisted of two types of customized multibranch modules exploiting multiscale feature representation to improve the robustness over local image noise and artifacts. Inserts with various densities of different materials [hydroxyapatite (HA), iodine, a blood–iodine mixture, and fat] were scanned using a research photon‐counting detector (PCD) CT with two energy thresholds and multiple radiation dose levels. The network was trained using phantom image patches only, and tested with different‐configurations of full field‐of‐view phantom and in vivo porcine images. Furthermore, the nominal mass densities of insert materials were used as the labels in CNN training, which potentially provided an implicit mass conservation constraint. The Incept‐net performance was evaluated in terms of image noise, detail preservation, and quantitative accuracy. Its performance was also compared to common material decomposition algorithms including least‐square‐based material decomposition (LS‐MD), total‐variation regularized material decomposition (TV‐MD), and U‐net‐based method.ResultsIncept‐net improved accuracy of the predicted mass density of basis materials compared with the U‐net, TV‐MD, and LS‐MD: the mean absolute error (MAE) of iodine was 0.66, 1.0, 1.33, and 1.57 mgI/cc for Incept‐net, U‐net, TV‐MD, and LS‐MD, respectively, across all iodine‐present inserts (2.0–24.0 mgI/cc). With the LS‐MD as the baseline, Incept‐net and U‐net achieved comparable noise reduction (both around 95%), both higher than TV‐MD (85%). The proposed IGC regularizer effectively helped both Incept‐net and U‐net to reduce image artifact. Incept‐net closely conserved the total mass densities (i.e., mass conservation constraint) in porcine images, which heuristically validated the quantitative accuracy of its outputs in anatomical background. In general, Incept‐net performance was less dependent on radiation dose levels than the two conventional methods; with approximately 40% less parameters, the Incept‐net achieved relatively improved performance than the comparator U‐net, indicating that performance gain by Incept‐net was not achieved by simply increasing network learning capacity.ConclusionIncept‐net demonstrated superior qualitative image appearance, quantitative accuracy, and lower noise than the conventional methods and less sensitive to dose change. Incept‐net generalized and performed well with unseen image structures and different material mass densities. This study provided preliminary evidence that the proposed CNN may be used to improve the material decomposition quality in multi‐energy CT.