AAGAN: Accuracy-Aware Generative Adversarial Network for Supervised Tasks

Many technical updates have been made in multi-detector CT. To evaluate image quality and radiation dose of high-pitch second- and third-generation dual-source chest CT angiography and to assess the effects of different levels of advanced modeled iterative reconstruction (ADMIRE) in newborns and children. Chest CT angiography (70 kVp) was performed in 42 children (age 158 ± 267 days, range 1-1,194 days). We evaluated subjective and objective image quality, and radiation dose with filtered back projection (FBP) and different strength levels of ADMIRE. For comparison were 42 matched controls examined with a second-generation 128-slice dual-source CT-scanner (80 kVp). ADMIRE demonstrated improved objective and subjective image quality (P < .01). Mean signal/noise, contrast/noise and subjective image quality were 11.9, 10.0 and 1.9, respectively, for the 80 kVp mode and 11.2, 10.0 and 1.9 for the 70 kVp mode. With ADMIRE, the corresponding values for the 70 kVp mode were 13.7, 12.1 and 1.4 at strength level 2 and 17.6, 15.6 and 1.2 at strength level 4. Mean CTDIvol, DLP and effective dose were significantly lower with the 70-kVp mode (0.31 mGy, 5.33 mGy*cm, 0.36 mSv) compared to the 80-kVp mode (0.46 mGy, 9.17 mGy*cm, 0.62 mSv; P < .01). The third-generation dual-source CT at 70 kVp provided good objective and subjective image quality at lower radiation exposure. ADMIRE improved objective and subjective image quality.

To identify the optimum strength of advanced modeled iterative reconstruction (ADMIRE) to achieve the best subjective and objective image quality when combining three-dose reduction strategies, ultra-high-pitch computed tomography coronary angiography (FLASH CTCA; with single-dose ivabradine to lower heart rate), low tube voltage, and ADMIRE. Sixty consecutive patients underwent FLASH CTCA at 100 kVp in this single-center prospective study. Single-dose ivabradine was administered to patients whose heart rate was above 75 bpm. Images were reconstructed using the three highest strengths of ADMIRE (A3, A4, and A5). Objective and subjective image quality (using a Likert scale) were evaluated in the three datasets. The signal strength remained unchanged but mean noise significantly reduced across the increasing strengths of ADMIRE [signal: 513.78 ± 101.7 Hounsfield units (HU) at A3, 515.6 ± 100.5 HU at A4, and 519.7 ± 107.9 HU at A5; noise: 23.4 ± 4.5 HU at A3, 20.2 ± 3.6 HU at A4, and 17.2 ± 3.3 HU at A5]. Signal-to-noise and contrast-to-noise ratios were the highest at A5, and A5 offered significantly higher Likert scores in image noise, vessel sharpness, and overall image quality than A3 or A4. Additionally, A5 did not interfere with image interpretation in any patient. Using all three dose reduction strategies during FLASH CTCA along with single-dose ivabradine administration ensures minimal radiation exposure in daily practice. In this study, A5 datasets had the best overall subjective and objective image quality despite their "plastic appearance". In the future, enhanced dose reduction can be obtained by further lowering tube voltages.

Evaluation of a Tube Voltage-Based Contrast Media Adaptation in Coronary Computed Tomography Angiography Using Personalized Triphasic Injection Protocols: A Matched Case-Control Study

Ultra-fast, low dose high-pitch (FLASH) versus prospectively-gated coronary computed tomography angiography: Comparison of image quality and patient radiation exposure

Magnetic Resonance Imaging (MRI) is one of the most commonly used modalities in medical imaging, and one of the main limitations is its slow scanning and reconstruction speed. Recently, deep learning used for compressed sensing (CS) methods have been proposed to accelerate the acquisition by undersampling in the K-space and reconstruct images with neural networks. However, there are still some challenges remained: First, directly training networks based on L1/L2 distance to the target fully sampled images may lead to fuzzy reconstruction images because L1/L2 loss only enforces the overall image or patch similarity, but does not consider the local details such as anatomical sharpness. Second, Generative Adversarial Networks (GAN) can partially solve this problem. The undersampling image gets the latent space through the encoder, and the image is reconstructed by the decoder based on GAN loss, but it may generate unrealistic details by lacking of constraints in K-space domain. Third, most of the networks after training are fixed and have limited adaptation capability in the inference time, and the patient-specific information cannot be effectively used. To resolve these challenges, we proposed a new compressed sensing GAN reconstruction method, and there are two main contributions: (1) We proposed a encoder-decoder structure, which guided GAN optimization strategy data-consistency in latent space to improve the reconstruction quality such as preserving more local details and improving the anatomical sharpness while constraining GAN to follow the data distribution in the K-space to prevent the unrealistic details. (2) An online update strategy is used to find the best representation in the latent space for the underlying patient, and the reconstruct result can be further improved by incorporating the patient-specific information. Extensive experimental results show the effectiveness of our method. KeywordsMRI reconstructionData-consistencyLatent spaceOnline update

Generative adversarial networks (GANs) have been increasingly used as feature mapping functions in speech enhancement, in which the noisy speech features are transformed to the clean ones through the generators. This article proposes a novel speech enhancement model based on a cycle-consistent relativistic GAN with Dilated Residual Networks and a Multi-attention mechanism. Using the adversarial loss, improved cycle-consistency losses, and an identity-mapping loss, a noisy-to-clean generator G and an inverse clean-to-noisy generator F simultaneously learn the forward and backward mappings between the source and target domains. To guarantee the stability of the training process, we replace vanilla GAN loss with relativistic average GAN loss and use spectral normalization in discriminators so that they conform to Lipschitz continuity. Furthermore, we employ two attention-based components as multi-attention mechanism to reduce importing signal distortion: attention U-net gates and dilated residual self-attention blocks. By employing these components, our proposed generators can capture long-term inner dependencies between elements of speech features and further preserve linguistic information. Experimental results on a public dataset indicate that the proposed model achieves state-of-the-art speech enhancement performance, especially in reducing speech distortion and improving signal overall quality. Compared with the representative GAN-based approaches, the proposed method significantly achieves the best performance in terms of STOI, CSIG, COVL, and CBAK objective metrics. Moreover, we demonstrate the contribution of each proposed component including relativistic average loss, attention U-net gate, self-attention layers, spectral normalization, and dilation operation by ten comparison systems.

Evaluating a low-dose protocol for computed tomography (CT) pulmonary angiography: combining reduced contrast medium, low tube voltage, and advanced image reconstruction.

Iterative reconstruction can be helpful to reduce radiation dose while maintaining image quality. However, this technique has not been fully evaluated in children during abdominal CT. To compare objective and subjective image quality between half-dose images reconstructed with iterative reconstruction at iteration strength levels 1 to 5 (half-S1 to half-S5 studies) and full-dose images reconstructed with filtered back projection (full studies) in pediatric abdominal CT. Twenty-one children (M:F = 13:8; mean age 8.2 ± 5.7years) underwent dual-source abdominal CT (mean effective dose 4.8 ± 2.1mSv). The objective image quality was evaluated as noise. Subjective image quality analysis was performed comparing each half study to the full study for noise, sharpness, artifact and diagnostic acceptability. Both objective and subjective image noise decreased with increasing iteration strength. Half-S4 and -S5 studies showed objective image noise similar to or lower than that of full studies. The half-S2 and -S3 studies produced the greatest sharpness and the half-S5 studies were the worst from a blocky appearance. Full and half studies did not differ in artifacts. Half-S3 studies showed the best diagnostic acceptability. Half-S4 and -S5 studies objectively and half-S3 studies subjectively showed comparable image quality to full studies in pediatric abdominal CT.

(1) Background: To evaluate the effects of an AI-based denoising post-processing software solution in low-dose whole-body computer tomography (WBCT) stagings; (2) Methods: From 1 January 2019 to 1 January 2021, we retrospectively included biometrically matching melanoma patients with clinically indicated WBCT staging from two scanners. The scans were reconstructed using weighted filtered back-projection (wFBP) and Advanced Modeled Iterative Reconstruction strength 2 (ADMIRE 2) at 100% and simulated 50%, 40%, and 30% radiation doses. Each dataset was post-processed using a novel denoising software solution. Five blinded radiologists independently scored subjective image quality twice with 6 weeks between readings. Inter-rater agreement and intra-rater reliability were determined with an intraclass correlation coefficient (ICC). An adequately corrected mixed-effects analysis was used to compare objective and subjective image quality. Multiple linear regression measured the contribution of “Radiation Dose”, “Scanner”, “Mode”, “Rater”, and “Timepoint” to image quality. Consistent regions of interest (ROI) measured noise for objective image quality; (3) Results: With good–excellent inter-rater agreement and intra-rater reliability (Timepoint 1: ICC ≥ 0.82, 95% CI 0.74–0.88; Timepoint 2: ICC ≥ 0.86, 95% CI 0.80–0.91; Timepoint 1 vs. 2: ICC ≥ 0.84, 95% CI 0.78–0.90; all p ≤ 0.001), subjective image quality deteriorated significantly below 100% for wFBP and ADMIRE 2 but remained good–excellent for the post-processed images, regardless of input (p ≤ 0.002). In regression analysis, significant increases in subjective image quality were only observed for higher radiation doses (≥0.78, 95%CI 0.63–0.93; p < 0.001), as well as for the post-processed images (≥2.88, 95%CI 2.72–3.03, p < 0.001). All post-processed images had significantly lower image noise than their standard counterparts (p < 0.001), with no differences between the post-processed images themselves. (4) Conclusions: The investigated AI post-processing software solution produces diagnostic images as low as 30% of the initial radiation dose (3.13 ± 0.75 mSv), regardless of scanner type or reconstruction method. Therefore, it might help limit patient radiation exposure, especially in the setting of repeated whole-body staging examinations.

Generative adversarial networks (GANs) have shown striking performances in computer vision applications to augment virtual training samples (VTS). However, the VTS generating by GANs in the context of hyperspectral image classification suffer from structural inconsistency due to the insufficient number of training samples in order to learn high-order features from the discriminator. This work addresses the scarcity of training samples by designing a GAN, in which the performance of discriminator is improved to produce more structurally coherent VTS. In the proposed method, by splitting the discriminator into two parts, GAN undertakes two tasks: the main task is to learn to distinguish between real and fake samples, and the auxiliary task is to learn to distinguish structurally corrupted and real samples. With this setup, GAN will produce real-like VTS with a higher variation than conventional GAN. Furthermore, in order to reduce the computational cost, subspace-based dimension reduction was performed to obtain the dominant features around the training samples to generate meaningful patterns from the original ones to be used in the learning phase. Based on the experimental results on real, and well-known hyperspectral benchmark images, the proposed method improves the performance compared with GANs-related, and conventional data augmentation strategies <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> .

The purpose of this study is to evaluate the radiation dose, image quality, and diagnostic accuracy of prospective ECG-gated cardiac CT at 70kV and adaptive statistical iterative reconstruction (ASIR), with a single source, 512 slice MDCT in the diagnosis of complex congenital heart disease in infants. We retrospectively evaluated 47 infants (ages 1day to 353days) with prospective ECG-gated cardiac CT that was performed on a single source 512 slice CT at low tube voltage (70kV) using a wide detector aperture, adaptive statistical iterative reconstruction algorithm (ASIR), and specific reconstruction software reducing coronary motion artifacts (SnapShot Freeze). All cardiac images were obtained during the first pass of contrast material through the anatomic structures of interest and the targets for the center of the acquisition window were set 45% of the R-R interval during one cardiac cycle without sedation and breath-hold. The median effective dose measured in our study was 0.64 ± 0.16mSv. The average subjective overall image quality score was 4.34 ± 0.31 (range 3-5). For the determination of objective image quality, Mean Noise (HU), SNR, and CNR values ​​emerged as 20.8, 28.7(for pulmonary artery), and 27.1, respectively. Diagnostic accuracy was 100% for the main purposes for the main clinical indication. During cardiac CT examination, pathologies in addition to cardiac anomalies were found in 9/47 of cases (7 severe airway obstructions,1 posterior diaphragmatic hernia, 1 vertebral anomaly).Prospective ECG-gated cardiac CT scan at 70kV and ASIR in infants with complex CHD provides low radiation dose (submillisievert) in a single heartbeat with a good objective and subjective image quality. It also provides important benefits in the diagnosis of additional pathology.

Generative Adversarial Network (GAN) is neural network architecture, widely used in many computer vision applications such as super-resolution image generation, art creation and image to image translation. A conventional GAN model consists of two sub-models; generative model and discriminative model. The former one generates new samples based on an unsupervised learning task, and the later one classifies them into real or fake. Though GAN is most commonly used for training generative models, it can be used for developing a classifier model. The main objective is to extend the effectiveness of GAN into semi-supervised learning, i.e., for the classification of fundus images to diagnose glaucoma. The discriminator model in the conventional GAN is improved via transfer learning to predict n + 1 classes by training the model for both supervised classification (n classes) and unsupervised classification (fake or real). Both models share all feature extraction layers and differ in the output layers. Thus any update in one of the model will impact both models. Results show that the semi-supervised GAN performs well than a standalone Convolution Neural Networks (CNNs) model.

We provide a general framework for designing Generative Adversarial Networks (GANs) to solve high-dimensional robust statistics problems, which aim at estimating unknown parameter of the true distribution given adversarially corrupted samples. Prior work [1], [2] focus on the problem of robust mean and covariance estimation when the true distribution lies in the family of Gaussian distributions or elliptical distributions, and analyze depth or scoring rule based GAN losses for the problem. Our work extend these to robust mean estimation, second-moment estimation, and robust linear regression when the true distribution only has bounded Orlicz norms, which includes the broad family of sub-Gaussian, sub-exponential and bounded moment distributions. We also provide a different set of sufficient conditions for the GAN loss to work: we only require its induced distance function to be a cumulative density function of some light-tailed distribution, which is easily satisfied by neural networks with sigmoid activation. In terms of techniques, our proposed GAN losses can be viewed as a smoothed and generalized Kolmogorov-Smirnov distance, which overcomes the computational intractability of the original Kolmogorov-Smirnov distance used in the weaken the distance approach in [3].

To compare image quality and radiation dose of abdominal split-filter dual-energy CT (SF-DECT) combined with monoenergetic imaging to single-energy CT (SECT) with automatic tube voltage selection (ATVS). Two-hundred single-source abdominal CT scans were performed as SECT with ATVS (n = 100) and SF-DECT (n = 100). SF-DECT scans were reconstructed and subdivided into composed images (SF-CI) and monoenergetic images at 55 keV (SF-MI). Objective and subjective image quality were compared among single-energy images (SEI), SF-CI and SF-MI. CNR and FOM were separately calculated for the liver (e.g. CNRliv) and the portal vein (CNRpv). Radiation dose was compared using size-specific dose estimate (SSDE). Results of the three groups were compared using non-parametric tests. Image noise of SF-CI was 18% lower compared to SEI and 48% lower compared to SF-MI (p < 0.001). Composed images yielded higher CNRliv over single-energy images (23.4 vs. 20.9; p < 0.001), whereas CNRpv was significantly lower (3.5 vs. 5.2; p < 0.001). Monoenergetic images overcame this inferiority in CNRpv and achieved similar results compared to single-energy images (5.1 vs. 5.2; p > 0.628). Subjective sharpness was equal between single-energy and monoenergetic images and diagnostic confidence was equal between single-energy and composed images. FOMliv was highest for SF-CI. FOMpv was equal for SEI and SF-MI (p = 0.78). SSDE was significant lower for SF-DECT compared to SECT (p < 0.022). The combined use of split-filter dual-energy CT images provides comparable objective and subjective image quality at lower radiation dose compared to single-energy CT with ATVS. • Split-filter dual-energy results in 18% lower noise compared to single-energy with ATVS. • Split-filter dual-energy results in 11% lower SSDE compared to single-energy with ATVS. • Spectral shaping of split-filter dual-energy leads to an increased dose-efficiency.

ObjectivesTo evaluate the diagnostic performance and image quality of an iterative model-based reconstruction (IMR) using a 100-kVp protocol for the assessment of heavily calcified coronary vessels, compared to those of filtered back projection (FBP) and hybrid iterative technique (iDose4), and also compared to those of IMR with standard 120 kVp protocol.MethodsAmong patients with Agatston scores ≥ 400 who had undergone both coronary CT angiography (CCTA) and invasive coronary angiography (ICA), age- and sex-matched patients with body mass index < 30 were retrospectively enrolled from CCTA with low-kVp protocol (100 kVp, n = 30) and with standard-kVp protocol (120 kVp, n = 30). Image data were all reconstructed with FBP, iDose4, and IMR. In each dataset, the objective and subjective image quality, and diagnostic accuracy (> 50% in luminal reduction as compared with ICA) were assessed.ResultsIMR showed better objective and subjective image quality than FBP and iDose4 in both 100 kVp and 120 kVp groups (all p < 0.05). IMR showed a significantly improved all diagnostic performance compared with FBP (p < 0.05). Compared with iDose4, IMR significantly improved positive predictive value (85.0% vs. 80.5%; p < 0.05). There was no significant difference in image quality and diagnostic performance using IMR between the 100 kVp and 120 kVp groups.Conclusions100 kVp IMR may be useful for the assessment of heavily calcified coronary vessels, providing better diagnostic performance than FBP or iDose4 at the same dose, while maintaining similar diagnostic accuracy to 120 kVp IMR.