3D printed phantom with 12 000 submillimeter lesions to improve efficiency in CT detectability assessment.

BACKGROUND: The detectability performance of a CT scanner is difficult to precisely quantify when nonlinearities are present in reconstruction. An efficient detectability assessment method that is sensitive to small effects of dose and scanner settings is desirable. We previously proposed a method using a search challenge instrument: a phantom is embedded with hundreds of lesions at random locations, and a model observer is used to detect lesions. Preliminary tests in simulation and a prototype showed promising results. PURPOSE: In this work, we fabricated a full-size search challenge phantom with design updates, including changes to lesion size, contrast, and number, and studied our implementation by comparing the lesion detectability from a nonprewhitening (NPW) model observer between different reconstructions at different exposure levels, and by estimating the instrument sensitivity to detect changes in dose. METHODS: Designed to fit into QRM anthropomorphic phantoms, our search challenge phantom is a cylindrical insert 10 cm wide and 4 cm thick, embedded with 12 000 lesions (nominal width of 0.6 mm, height of 0.8 mm, and contrast of -350 HU), and was fabricated using PixelPrint, a 3D printing technique. The insert was scanned alone at a high dose to assess printing accuracy. To evaluate lesion detectability, the insert was placed in a QRM thorax phantom and scanned from 50 to 625 mAs with increments of 25 mAs, once per exposure level, and the average of all exposure levels was used as high-dose reference. Scans were reconstructed with three different settings: filtered-backprojection (FBP) with Br40 and Br59, and Sinogram Affirmed Iterative Reconstruction (SAFIRE) with strength level 5 and Br59 kernel. An NPW model observer was used to search for lesions, and detection performance of different settings were compared using area under the exponential transform of free response ROC curve (AUC). Using propagation of uncertainty, the sensitivity to changes in dose was estimated by the percent change in exposure due to one standard deviation of AUC, measured from 5 repeat scans at 100, 200, 300, and 400 mAs. RESULTS: The printed insert lesions had an average position error of 0.20 mm compared to printing reference. As the exposure level increases from 50 mAs to 625 mAs, the lesion detectability AUCs increase from 0.38 to 0.92, 0.42 to 0.98, and 0.41 to 0.97 for FBP Br40, FBP Br59, and SAFIRE Br59, respectively, with a lower rate of increase at higher exposure level. FBP Br59 performed best with AUC 0.01 higher than SAFIRE Br59 on average and 0.07 higher than FBP Br40 (all P < 0.001). The standard deviation of AUC was less than 0.006, and the sensitivity to detect changes in mAs was within 2% for FBP Br59. CONCLUSIONS: Our 3D-printed search challenge phantom with 12 000 submillimeter lesions, together with an NPW model observer, provide an efficient CT detectability assessment method that is sensitive to subtle effects in reconstruction and is sensitive to small changes in dose.

Evaluation of data uncertainty for deep-learning-based CT noise reduction using ensemble patient data and a virtual imaging trial framework.

Deep learning-based image reconstruction and noise reduction (DLIR) methods have been increasingly deployed in clinical CT. Accurate assessment of their data uncertainty properties is essential to understand the stability of DLIR in response to noise. In this work, we aim to evaluate the data uncertainty of a DLIR method using real patient data and a virtual imaging trial framework and compare it with filtered-backprojection (FBP) and iterative reconstruction (IR). The ensemble of noise realizations was generated by using a realistic projection domain noise insertion technique. The impact of varying dose levels and denoising strengths were investigated for a ResNet-based deep convolutional neural network (DCNN) model trained using patient images. On the uncertainty maps, DCNN shows more detailed structures than IR although its bias map has less structural dependency, which implies that DCNN is more sensitive to small changes in the input. Both visual examples and histogram analysis demonstrated that hotspots of uncertainty in DCNN may be associated with a higher chance of distortion from the truth than IR, but it may also correspond to a better detection performance for some of the small structures.

Automated Web-based Software for CT Quality Control Testing of Low-contrast Detectability using Model Observers.

The Channelized Hotelling observer (CHO) is well correlated with human observer performance in many CT detection/classification tasks but has not been widely adopted in routine CT quality control and performance evaluation, mainly because of the lack of an easily available, efficient, and validated software tool. We developed a highly automated solution - CT image quality evaluation and Protocol Optimization (CTPro), a web-based software platform that includes CHO and other traditional image quality assessment tools such as modulation transfer function and noise power spectrum. This tool can allow easy access to the CHO for both the research and clinical community and enable efficient, accurate image quality evaluation without the need of installing additional software. Its application was demonstrated by comparing the low-contrast detectability on a clinical photon-counting-detector (PCD)-CT with a traditional energy-integrating-detector (EID)-CT, which showed UHR-T3D had 6.2% higher d' than EID-CT with IR (p = 0.047) and 4.1% lower d' without IR (p = 0.122).

Image quality evaluation in deep-learning-based CT noise reduction using virtual imaging trial methods: Contrast-dependent spatial resolution.

BACKGROUND: Deep-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. PURPOSE: In 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. METHODS: The 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. RESULTS: The 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. CONCLUSIONS: A 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.

Spectral information content of Compton scattering events in silicon photon counting detectors.

BACKGROUND: Silicon (Si) is a possible sensor material for photon counting detectors (PCDs). A major drawback of Si is that roughly two-thirds of x-ray interactions in the diagnostic energy range are Compton scattering. Because Compton scattering is an energy-insensitive process, it is commonly assumed that Compton events retain little spectral information. PURPOSE: To quantify how much information can be recovered from Compton scattering events in models of Si PCDs. METHODS: We built a simplified model of Si interactions including two interaction mechanisms: photoelectric effect and Compton scattering. We considered three different binning options that represent strategies for handling Compton events: in Compton censoring, all events under 38 keV (the maximum energy possible from Compton scattering for a 120 keV incident photon) were discarded; in Compton counting, all events between 1 and 38 keV were placed into a single bin; in Compton binning, all events were placed into energy bins of uniform width. These were compared to the ideal detector, which always recorded the correct energy (i.e., 100% photoelectric effect). Every photon was assumed to interact once and only once with Si, and the energy bin width was 5 keV. In the primary analysis, the Si detector was irradiated with a 120 kV spectrum filtered by 30 cm of water, with 99.5% of the arriving spectrum above 38 keV so that there was good separation between photoelectric effect and Compton scattering, and the figures of merit were the Cramér-Rao lower bound (CRLB) of the variance of iodine and water basis material decomposition images, as well as the CRLB of virtual monoenergetic images (i.e., linear combinations of material images) that maximize iodine CNR or water CNR. We also constructed a local linear estimator that attains the CRLB. In secondary analyses, we applied other sources of spectral distortion: (1) a nonzero minimum energy threshold; (2) coarser, 10 keV energy bins; and (3) a model of charge sharing. RESULTS: With our chosen spectrum, 67% of the interactions were Compton scattering. Consistent with this, the material decomposition variance for the Compton censoring model, averaged over both basis materials, was 258% greater than the ideal detector. If Compton events carried no spectral information, the Compton counting model would show similar variance. Instead, its basis material variance was 103% greater than the ideal detector, implying that Compton counts indeed carry significant spectral information. The Compton binning model had a basis material variance 60% greater than the ideal detector. The Compton binning model was not affected by a 5 keV minimum energy threshold, but the variance increased from 60% to 107% when charge sharing was included and to 78% with coarser energy bins. For optimized CNR images, the average variance was 149%, 12%, and 10% higher than the ideal detector for the Compton censoring, counting, and binning models, reinforcing the hypothesis that Compton counts are useful for detection tasks and that precise energy assignments are not necessary. CONCLUSIONS: Substantial spectral information remains after Compton scattering events in silicon PCDs.

Possible improvements in effective fill factor using X-ray fluorescent interpixel reflectors.

BACKGROUND: The spatial resolution of energy-integrating diagnostic CT scanners is limited by interpixel reflectors on the detector, which optically isolate pixels but create dead space. Because the width of the reflector cannot easily be decreased, fill factor diminishes as resolution increases. PURPOSE: We propose loading (or mixing) a high-Z element into the reflectors, causing the reflectors to be X-ray fluorescent. Re-emitted characteristic X-rays could be detected in adjacent pixels, increasing the effective fill factor and compensating for fill factor loss with higher-resolution detectors. The purpose of this work is to understand the physical principles of this approach and to analyze its effectiveness using Monte Carlo simulations. METHODS: Detector pixels were modeled using the GEANT4 Monte Carlo package. The width of the reflector was kept constant at 0.1 mm throughout, and we considered pixel pitches between 0.5 and 1 mm. The pixelated scintillator material was gadolinium oxysulfide, 3 mm thick. The baseline reflector material was chosen to be acrylic, and varying concentrations of a high-Z element were loaded into the material. We assumed that the optical characteristics of pixels were ideal (no absorption within pixels, perfect reflection at boundaries). The detector was irradiated uniformly with 10,000 X-ray photons to estimate its spectral response. The figure of merit was the variance of the detector signal at zero frequency normalized to that of an ideal single-bin photon-counting detector with 100% fill factor. Sensitivity analyses were conducted to understand the effect of varying the high-Z element concentration and the spectrum. RESULTS: Initial simulations suggested that a k-edge near 50 keV would be ideal. Gd was therefore selected as the high-Z material. The relative variances for a conventional energy integrating detector without Gd at 1 mm pixel pitch (81% fill factor) and 0.5 mm pixel pitch (64% fill factor) were 1.38 and 1.74, compared to 1.00 for an ideal photon counting detector, implying a 26% variance penalty for 0.5 mm pitch. When 1 g/cm3 Gd was loaded into the interpixel reflector, the relative variance improved to 1.27 and 1.43, respectively, implying that the variance penalty for including Gd together with 0.5 mm pitch is only 4%. Performance was nearly maximized at 1.0 g/cm3 of Gd, but a concentration of 0.5 g/cm3 of Gd showed most of the benefit. Improvements depend weakly on kV, with lower kV associated with higher improvements. An external anti-scatter grid was not modeled in our simulations and would reduce the expected benefit, depending greatly on the pitch and dimensionality of the anti-scatter grid. CONCLUSIONS: The losses in fill factor associated with smaller pixel pitch can be reduced if Gd or a similar element could be loaded into the interpixel reflector. These improvements in noise efficiency are yet to be verified experimentally.

Targeted Training Reduces Search Errors but Not Classification Errors for Hepatic Metastasis Detection at Contrast-Enhanced CT.

RATIONALE AND OBJECTIVES: Methods are needed to improve the detection of hepatic metastases. Errors occur in both lesion detection (search) and decisions of benign versus malignant (classification). Our purpose was to evaluate a training program to reduce search errors and classification errors in the detection of hepatic metastases in contrast-enhanced abdominal computed tomography (CT). MATERIALS AND METHODS: After Institutional Review Board approval, we conducted a single-group prospective pretest-posttest study. Pretest and posttest were identical and consisted of interpreting 40 contrast-enhanced abdominal CT exams containing 91 liver metastases under eye tracking. Between pretest and posttest, readers completed search training with eye-tracker feedback and coaching to increase interpretation time, use liver windows, and use coronal reformations. They also completed classification training with part-task practice, rating lesions as benign or malignant. The primary outcome was metastases missed due to search errors (<2 seconds gaze under eye tracker) and classification errors (>2 seconds). Jackknife free-response receiver operator characteristic (JAFROC) analysis was also conducted. RESULTS: A total of 31 radiologist readers (8 abdominal subspecialists, 8 nonabdominal subspecialists, 15 senior residents/fellows) participated. Search errors were reduced (pretest 11%, posttest 8%, difference 3% [95% confidence interval, 0.3%-5.1%], P = .01), but there was no difference in classification errors (difference 0%, P = .97) or in JAFROC figure of merit (difference -0.01, P = .36). In subgroup analysis, abdominal subspecialists demonstrated no evidence of change. CONCLUSION: Targeted training reduced search errors but not classification errors for the detection of hepatic metastases at contrast-enhanced abdominal CT. Improvements were not seen in all subgroups.

Direct energy binning for photon counting detectors: Simulation study.

BACKGROUND: Photon counting detectors (PCDs) for x-ray computed tomography (CT) face spectral distortion from pulse pileup and charge sharing. The photon counting scheme used by many PCDs is threshold-subtract (TS) with pulse height analysis (PHA), where each counter counts up-crossing events when pulses exceed an energy threshold. PCD data are not Poisson-distributed due to charge sharing and pulse pileup, but the counting statistics have never been studied yet. PURPOSE: The objectives of this study were (1) to propose a modified photon counting scheme, direct energy binning (DB), that is expected to be robust against pulse pileup; (2) to assess the performance of DB compared to TS; and (3) to evaluate its counting statistics. METHODS: With DB scheme, counter k starts a timer upon an up-crossing event of energy threshold k, and adds a count only if the next higher energy threshold (k+1) was not crossed within a short time window (hence, the pulse peak belongs to the energy bin k). We used Monte Carlo (MC) simulation and assessed count-rate curves and count-rate-dependent spectral imaging task performance for conventional CT imaging as well as water thickness estimation, water-bone material decomposition, and K-edge imaging with tungsten as the K-edge material. We also assessed count-rate-dependent measurement statistics such as expectation, variance, and covariance of total counts as well as energy bin outputs. The agreement with counting statistics models was also evaluated. RESULTS: The DB scheme improved the count-rate curve, that is, mean measured counts as a function of input count-rate, and peaked with 59% higher count-rate capability than the TS scheme (3.5 × 108 counts per second (cps)/mm2 versus 2.3 × 108  cps/mm2 ). The Cramér-Rao lower bounds (CRLB) of the variance of basis line integrals estimation for DB was better than those for TS by 2% for the conventional CT imaging, 30% for water-bone material decomposition, and 32% for K-edge imaging at 1000 mA (at 7.3 × 107  cps/sub-pixel after charge sharing). When count-rates were lower, PCD data statistics were dominated by charge sharing: the variance of total counts and lower energy bins was larger than the mean counts; the covariance of bin data was positive and non-zero. When count-rates were higher, PCD data statistics were dominated by pulse pileup: the variance of data was lower than the mean; the covariance of bin data was negative. The transition between the two regimes occurred smoothly, and pulse pileup dominated the statistics ≥400 mA (when the count-rate after charge sharing was 2.9 × 107  cps/sub-pixel and the probability of count-loss for DB was 37%). Both DB and TS had good agreement with Yu-Fessler's models of total counts; however, DB had a better agreement with Wang's variance and covariance models for energy bin data than TS did. CONCLUSIONS: The proposed DB scheme had several advantages over TS. At low to moderate flux, DB could improve the resilience of PCDs to pulse pileup. Counting statistics deviated from the Poisson distribution due to charge sharing for lower count-rate conditions and pulse pileup for higher count-rate conditions.

Supervised Learning for CT Denoising and Deconvolution Without High-Resolution Reference Images.

Purpose: Convolutional neural networks (CNNs) have been proposed for super-resolution in CT, but training of CNNs requires high-resolution reference data. Higher spatial resolution can also be achieved using deconvolution, but conventional deconvolution approaches amplify noise. We develop a CNN that mitigates increasing noise and that does not require higher-resolution reference images. Methods: Our model includes a noise reduction CNN and a deconvolution CNN that are separately trained. The noise reduction CNN is a U-Net, similar to other noise reduction CNNs found in the literature. The deconvolution CNN uses an autoencoder, where the decoder is fixed and provided as a hyperparameter that represents the system point spread function. The encoder is trained to provide a deconvolution that does not amplify noise. Ringing can occur from deconvolution but is controlled with a difference of gradients loss function term. Our technique was demonstrated on a variety of patient images and on ex vivo kidney stones. Results: The noise reduction and deconvolution CNNs produced visually sharper images at low noise. In ex vivo mixed kidney stones, better visual delineation of the kidney stone components could be seen. Conclusions: A noise reduction and deconvolution CNN improves spatial resolution and reduces noise without requiring higher-resolution reference images.

Technical note: Exploring the detectability of coronary calcification using ultra-high-resolution photon-counting-detector CT.

BACKGROUND: Coronary calcification is a strong indicator of coronary artery disease, and patients with a "zero" coronary calcification score have a much lower risk of future cardiac events than those with even small amounts of calcium. However, false-negative (incorrect zero scores) may occur if small calcifications are missed at CT due to limited spatial resolution. PURPOSE: To demonstrate lower limits of detection for coronary calcification using an ultra-high-resolution (UHR) mode on a clinical photon-counting-detector CT (PCD-CT), compared to a conventional energy-integrating-detector CT (EID-CT). METHODS: Chicken eggshell fragments (0.4-0.8 mm) mimicking coronary calcifications were scanned on a clinical PCD-CT (NAEOTOM Alpha) in UHR mode and a conventional EID-CT (SOMATOM Force) with matched tube potential and radiation dose levels to the PCD-CT. PCD-CT images were reconstructed with a sharp kernel (Qr68) and a quantum iterative algorithm (QIR-3). Two sets of EID-CT images were reconstructed: routine clinical kernel (Qr36, ADMIRE-3) and a sharper kernel (Qr54) with similar noise to PCD-CT images. With institutional review board approval, in vivo exams performed with the PCD-CT in UHR mode were compared against patients' clinical EID-CT exams. The visibility of calcifications on PCD-CT and EID-CT images was assessed and compared qualitatively. RESULTS: PCD-CT images visualized all calcified fragments, while EID-CT failed to detect those below 0.6 mm using a routine protocol. EID-CT with Qr54 improved visibility but distorted boundaries. Calcifications were less visible on EID-CT than PCD-CT as phantom sizes increased. 0.6- and 0.7-mm calcified fragments were barely visible on 35- and 40-cm phantom EID-CT images. Patient cases showed small calcifications missed on EID-CT but detected on PCD-CT. CONCLUSION: At matched radiation dose, PCD-CT in UHR mode provided higher spatial resolution and improved the detectability of small calcified fragments for different phantom/patient sizes in comparison to EID-CT.

Improvements in dose efficiency with high resolution scan modes in photon counting CT.

For the detection of very small objects, high resolution detectors are expected to provide higher dose efficiency. We assessed this impact of increased resolution on a clinical photon counting detector CT (PCD-CT) by comparing its detectability in high resolution and standard resolution (with 2×2 binning and larger focal spot) modes. A 50µm-thin metal wire was placed in a thorax phantom and scanned in both modes at three exposure levels (12, 15, and 18 mAs); acquired data were reconstructed with three reconstruction kernels (Br40, Br68, and Br76, from smooth to sharp). A scanning nonprewhitening model observer searched for the wire location within each slice independently. Detection performance was quantified as area under the exponential transform of the free response ROC curve. The high-resolution mode had the mean AUCs at 18 mAs of 0.45, 0.49, and 0.65 for Br40, Br68, and Br76, respectively, which were 2 times, 3.6 times, and 4.6 times those of the standard resolution mode. The high-resolution mode achieved greater AUC at 12 mAs than the standard resolution mode at 18 mAs for every reconstruction kernel, but improvements were larger at sharper kernels. The results are consistent with the greater suppression of noise aliasing expected at higher frequencies with high resolution CT. This work illustrates that PCD-CT can provide large dose efficiency gains for detection tasks of small, high contrast lesions.

Contrast- and noise-dependent spatial resolution measurement for deep convolutional neural network-based noise reduction in CT using patient data.

Deep convolutional neural network (DCNN)-based noise reduction methods have been increasingly deployed in clinical CT. Accurate assessment of their spatial resolution properties is required. Spatial resolution is typically measured on physical phantoms, which may not represent the true performance of DCNN in patients as it is typically trained and tested with patient images and the generalizability of DNN to physical phantoms is questionable. In this work, we proposed a patient-data-based framework to measure the spatial resolution of DCNN methods, which involves lesion- and noise-insertion in projection domain, lesion ensemble averaging, and modulation transfer function measurement using an oversampled edge spread function from the cylindrical lesion signal. The impact of varying lesion contrast, dose levels, and CNN denoising strengths were investigated for a ResNet-based DCNN model trained using patient images. The spatial resolution degradation of DCNN reconstructions becomes more severe as the contrast or radiation dose decreased, or DCNN denoising strength increased. The measured 50%/10% MTF spatial frequencies of DCNN with highest denoising strength were (-500 HU:0.36/0.72 mm-1; -100 HU:0.32/0.65 mm-1; -50 HU:0.27/0.53 mm-1; -20 HU:0.18/0.36 mm-1; -10 HU:0.15/0.30 mm-1), while the 50%/10% MTF values of FBP were almost kept constant of 0.38/0.76 mm-1.

Patient-specific uncertainty and bias quantification of non-transparent convolutional neural network model through knowledge distillation and Bayesian deep learning.

Assessing the reliability of convolutional neural network (CNN)-based CT imaging techniques is critical for reliable deployment in practice. Some evaluation methods exist but require full access to target CNN architecture and training data, something not available for proprietary or commercial algorithms. Moreover, there is a lack of systematic evaluation methods. To address these issues, we propose a patient-specific uncertainty and bias quantification (UNIQ) method that integrates knowledge distillation and Bayesian deep learning. Knowledge distillation creates a transparent CNN ("Student CNN") to approximate the target non-transparent CNN ("Teacher CNN"). Student CNN is built as a Bayesian-deep-learning-based probabilistic CNN that, for each input, always generates statistical distribution of the corresponding outputs, and characterizes predictive mean and two major uncertainties - data and model uncertainty. UNIQ was evaluated using a low-dose CT denoising task. Patient and phantom scans with routine-dose and synthetic quarter-dose were used to create training, validation, and testing sets. To demonstrate, Unet and Resnet were used as backbones of Teacher CNN and Student CNN respectively and were trained using independent training sets. Student Resnet was qualitatively and quantitatively evaluated. The pixel-wise predictive mean, data uncertainty, and model uncertainty from Student Resnet were very similar to the counterparts from Teacher Unet (mean-absolute-error: predictive mean 1.5HU, data uncertainty 1.8HU, model uncertainty 1.3HU; mean 2D correlation coefficient: total uncertainty 0.90, data uncertainty 0.86, model uncertainty 0.83). The proposed UNIQ can potentially systematically characterize the reliability of non-transparent CNN models used in CT.

A training program to reduce reader search errors for liver metastasis detection in CT.

Detection of low contrast liver metastases varies between radiologists. Training may improve performance for lower-performing readers and reduce inter-radiologist variability. We recruited 31 radiologists (15 trainees, 8 non-abdominal staff, and 8 abdominal staff) to participate in four separate reading sessions: pre-test, search training, classification training, and post-test. In the pre-test, each radiologist interpreted 40 liver CT exams containing 91 metastases, circumscribed suspected hepatic metastases while under eye tracker observation, and rated confidence. In search training, radiologists interpreted a separate set of 30 liver CT exams while receiving eye tracker feedback and after coaching to increase use of coronal reformations, interpretation time, and use of liver windows. In classification training, radiologists interpreted up to 100 liver CT image patches, most with benign or malignant lesions, and compared their annotations to ground truth. Post-test was identical to pre-test. Between pre- and post-test, sensitivity increased by 2.8% (p = 0.01) but AUC did not change significantly. Missed metastases were classified as search errors (<2 seconds gaze time) or classification errors (>2 seconds gaze time) using the eye tracker. Out of 2775 possible detections, search errors decreased (10.8% to 8.1%; p < 0.01) but classification errors were unchanged (5.7% vs 5.7%). When stratified by difficulty, easier metastases showed larger reductions in search errors: for metastases with average sensitivity of 0-50%, 50-90%, and 90-100%, reductions in search errors were 16%, 35%, and 58%, respectively. The training program studied here may be able to improve radiologist performance by reducing errors but not classification errors.

Classification of high-risk coronary plaques using radiomic analysis of multi-energy photon-counting-detector computed tomography (PCD-CT) images.

Coronary plaque risk classification in images acquired with photon-counting-detector (PCD) CT was performed using a radiomics-based machine learning (ML) model. With IRB approval, 19 coronary CTA patients were scanned on a PCD-CT (NAEOTOM Alpha, Siemens Healthineers) with median CTDIvol of 8.02 mGy. Five types of images: virtual monoenergetic images (VMIs) at 50-keV, 70-keV, and 100-keV, iodine maps, and virtual non-contrast (VNC) images were reconstructed using an iterative reconstruction algorithm (QIR), a quantitative kernel (Qr40) and 0.6-mm/0.3-mm slice thickness/increment. Atherosclerotic plaques were segmented using semi-automatic software (Research Frontier, Siemens). Segmentation confirmation and risk stratification (low- vs high-risk) were performed by a board-certified cardiac radiologist. A total of 93 radiomic features were extracted from each image using PyRadiomics (v2.2.0b1). For each feature, a t-test was performed between low- and high-risk plaques (p<0.05 considered significant). Two significant and non-redundant features were input into a support vector machine (SVM). A leave-one-out cross-validation strategy was adopted and the classification accuracy was computed. Fifteen low-risk and ten high-risk plaques were identified by the radiologist. A total of 18, 32, 43, 16, and 55 out of 93 features in 50-keV, 70-keV, 100-keV, iodine map, and VNC images were statistically significant. A total of 17, 19, 22, 20, and 22 out of 25 plaques were classified correctly in 50-keV, 70-keV, 100-keV, iodine map, and VNC images, respectively. A ML model using 100-keV VMIs and VNC images derived from coronary PCD-CTA best automatically differentiated low- and high-risk coronary plaques.

Aberration correction in diagnostic ultrasound: A review of the prior field and current directions.

Medical ultrasound images are reconstructed with simplifying assumptions on wave propagation, with one of the most prominent assumptions being that the imaging medium is composed of a constant sound speed. When the assumption of a constant sound speed are violated, which is true in most in vivoor clinical imaging scenarios, distortion of the transmitted and received ultrasound wavefronts appear and degrade the image quality. This distortion is known as aberration, and the techniques used to correct for the distortion are known as aberration correction techniques. Several models have been proposed to understand and correct for aberration. In this review paper, aberration and aberration correction are explored from the early models and correction techniques, including the near-field phase screen model and its associated correction techniques such as nearest-neighbor cross-correlation, to more recent models and correction techniques that incorporate spatially varying aberration and diffractive effects, such as models and techniques that rely on the estimation of the sound speed distribution in the imaging medium. In addition to historical models, future directions of ultrasound aberration correction are proposed.

In Vivo Prediction of Kidney Stone Fragility Using Radiomics-Based Regression Models.

Introduction: The surgical technique for urinary stone removal is partly influenced by its fragility, as prognosticated by the clinician. This feasibility study aims to develop a linear regression model from CT-based radiomic markers to predict kidney stone comminution time in vivo with two ultrasonic lithotrites. Materials and Methods: Patients identified by urologists at our institution as eligible candidates for percutaneous nephrolithotomy were prospectively enrolled. The active engagement time of the lithotrite in breaking the stone during surgery denoted the comminution time of each stone. The comminution rate was computed as the stone volume disintegrated per minute. Stones were grouped into three fragility classes (fragile, moderate, hard), based on inverse of the comminution rates with respect to the mean. Multivariable linear regression models were trained with radiomic features extracted from clinical CT images to predict comminution times in vivo. The model with the least root mean squared error (RMSE) on comminution times and the fewest misclassification of fragility was finally selected. Results: Twenty-eight patients with 31 stones in total were included in this study. Stones in the cohort averaged 1557 (±2472) mm3 in volume and 5.3 (±7.4) minutes in comminution time. Ten stones had nonmoderate fragility. Linear regression of stone volume alone predicted comminution time with an RMSE of 6.8 minutes and missed all 10 stones with nonmoderate fragility. A fragility model that included stone volume, internal morphology, shape-based radiomics, and device type improved RMSE to below 3.3 minutes and correctly classified 20/21 moderate and 6/10 nonmoderate stones. Conclusions: CT metrics-based fragility models may provide information to surgeons regarding kidney stone fragility and facilitate the selection of stone removal procedures.

Understanding Reader Variability: A 25-Radiologist Study on Liver Metastasis Detection at CT.

Background Substantial interreader variability exists for common tasks in CT imaging, such as detection of hepatic metastases. This variability can undermine patient care by leading to misdiagnosis. Purpose To determine the impact of interreader variability associated with (a) reader experience, (b) image navigation patterns (eg, eye movements, workstation interactions), and (c) eye gaze time at missed liver metastases on contrast-enhanced abdominal CT images. Materials and Methods In a single-center prospective observational trial at an academic institution between December 2020 and February 2021, readers were recruited to examine 40 contrast-enhanced abdominal CT studies (eight normal, 32 containing 91 liver metastases). Readers circumscribed hepatic metastases and reported confidence. The workstation tracked image navigation and eye movements. Performance was quantified by using the area under the jackknife alternative free-response receiver operator characteristic (JAFROC-1) curve and per-metastasis sensitivity and was associated with reader experience and image navigation variables. Differences in area under JAFROC curve were assessed with the Kruskal-Wallis test followed by the Dunn test, and effects of image navigation were assessed by using the Wilcoxon signed-rank test. Results Twenty-five readers (median age, 38 years; IQR, 31-45 years; 19 men) were recruited and included nine subspecialized abdominal radiologists, five nonabdominal staff radiologists, and 11 senior residents or fellows. Reader experience explained differences in area under the JAFROC curve, with abdominal radiologists demonstrating greater area under the JAFROC curve (mean, 0.77; 95% CI: 0.75, 0.79) than trainees (mean, 0.71; 95% CI: 0.69, 0.73) (P = .02) or nonabdominal subspecialists (mean, 0.69; 95% CI: 0.60, 0.78) (P = .03). Sensitivity was similar within the reader experience groups (P = .96). Image navigation variables that were associated with higher sensitivity included longer interpretation time (P = .003) and greater use of coronal images (P < .001). The eye gaze time was at least 0.5 and 2.0 seconds for 71% (266 of 377) and 40% (149 of 377) of missed metastases, respectively. Conclusion Abdominal radiologists demonstrated better discrimination for the detection of liver metastases on abdominal contrast-enhanced CT images. Missed metastases frequently received at least a brief eye gaze. Higher sensitivity was associated with longer interpretation time and greater use of liver display windows and coronal images. © RSNA, 2022 Online supplemental material is available for this article.

Individualized and generalized models for predicting observer performance on liver metastasis detection using CT.

Purpose: Radiologists exhibit wide inter-reader variability in diagnostic performance. This work aimed to compare different feature sets to predict if a radiologist could detect a specific liver metastasis in contrast-enhanced computed tomography (CT) images and to evaluate possible improvements in individualizing models to specific radiologists. Approach: Abdominal CT images from 102 patients, including 124 liver metastases in 51 patients were reconstructed at five different kernels/doses using projection domain noise insertion to yield 510 image sets. Ten abdominal radiologists marked suspected metastases in all image sets. Potentially salient features predicting metastasis detection were identified in three ways: (i) logistic regression based on human annotations (semantic), (ii) random forests based on radiologic features (radiomic), and (iii) inductive derivation using convolutional neural networks (CNN). For all three approaches, generalized models were trained using metastases that were detected by at least two radiologists. Conversely, individualized models were trained using each radiologist's markings to predict reader-specific metastases detection. Results: In fivefold cross-validation, both individualized and generalized CNN models achieved higher area under the receiver operating characteristic curves (AUCs) than semantic and radiomic models in predicting reader-specific metastases detection ability ( p < 0.001 ). The individualized CNN with an AUC of mean (SD) 0.85(0.04) outperformed the generalized one [ AUC = 0.78 ( 0.06 ) , p = 0.004 ]. The individualized semantic [ AUC = 0.70 ( 0.05 ) ] and radiomic models [ AUC = 0.68 ( 0.06 ) ] outperformed the respective generalized versions [semantic AUC = 0.66 ( 0.03 ) , p = 0.009 ; radiomic AUC = 0.64 ( 0.06 ) , p = 0.03 ]. Conclusions: Individualized models slightly outperformed generalized models for all three feature sets. Inductive CNNs were better at predicting metastases detection than semantic or radiomic features. Generalized models have implementation advantages when individualized data are unavailable.

Individualized and Generalized Learner Models for Predicting Missed Hepatic Metastases.

The diagnostic performance of radiologist readers exhibits substantial variation that cannot be explained by CT acquisition protocol differences. Studying reader detectability from CT images may help identify why certain types of lesions are missed by multiple or specific readers. Ten subspecialized abdominal radiologists marked all suspected metastases in a multi-reader-multi-case study of 102 deidentified contrast-enhanced CT liver scans at multiple radiation dose levels. A reference reader marked ground truth metastatic and benign lesions with the aid of histopathology or tumor progression on later scans. Multi-slice image patches and 3D radiomic features were extracted from the CT images. We trained deep convolutional neural networks (CNN) to predict whether an average (generalized) or individual radiologist reader would detect or miss a specific metastasis from an image patch containing it. The individualized CNN showed higher performance with an area under the receiver operating characteristic curve (AUC) of 0.82 compared to a generalized one (AUC = 0.78) in predicting reader-specific detectability. Random forests were used to build the respective versions from radiomic features. Both the individualized (AUC = 0.64) and generalized (AUC = 0.59) predictors from radiomic features showed limited ability to differentiate detected from missed lesions. This shows that CNN can identify and learn automated features that are better predictors of reader detectability of lesions than radiomic features. Individualized prediction of difficult lesions may allow targeted training of idiosyncratic weaknesses but requires substantial training data for each reader.