Medical Image Synthesis with Deep Convolutional Adversarial Networks

Detecting brain lesions in suspected acute ischemic stroke with CT-based synthetic MRI using generative adversarial networks

Background

Difficulties in detecting brain lesions in acute ischemic stroke (AIS) have convinced researchers to use computed tomography (CT) to scan for and magnetic resonance imaging (MRI) to search for these lesions. This work aimed to develop a generative adversarial network (GAN) model for CT-to-MR image synthesis and evaluate reader performance with synthetic MRI (syn-MRI) in detecting brain lesions in suspected patients.

Methods

Patients with primarily suspected AIS were randomly assigned to the training (n=140) or testing (n=53) set. Emergency CT and follow-up MR images in the training set were used to develop a GAN model to generate syn-MR images from the CT data in the testing set. The standard reference was the manual segmentations of follow-up MR images. Image similarity was evaluated between syn-MRI and the ground truth using a 4-grade visual rating scale, the peak signal-to-noise ratio (PSNR), and the structural similarity index measure (SSIM). Reader performance with syn-MRI and CT was evaluated and compared on a per-patient (patient detection) and per-lesion (lesion detection) basis. Paired t-tests or Wilcoxon signed-rank tests were used to compare reader performance in lesion detection between the syn-MRI and CT data.

Results

Grade 2–4 brain lesions were observed on syn-MRI in 92.5% (49/53) of the patients, while the remaining syn-MRI data showed no lesions compared to the ground truth. The GAN model exhibited a weak PSNR of 24.30 dB but a favorable SSIM of 0.857. Compared with CT, syn-MRI led to an increase in the overall sensitivity from 38% (57/150) to 82% (123/150) in patient detection and from 4% (68/1,620) to 16% (262/1,620) in lesion detection (R=0.32, corrected P<0.001), but the specificity in patient detection decreased from 67% (6/9) to 33% (3/9). An additional 75% (70/93) of patients and 15% (77/517) of lesions missed on CT were detected on syn-MRI.

Conclusions

The GAN model holds potential for generating synthetic MR images from noncontrast CT data and thus could help sensitively detect individuals among patients with suspected AIS. However, the image similarity performance of the model needs to be improved, and further expert discrimination is strongly recommended.

Synthesizing high-resolution magnetic resonance imaging using parallel cycle-consistent generative adversarial networks for fast magnetic resonance imaging

PURPOSE

The common practice in acquiring the magnetic resonance (MR) images is to obtain two-dimensional (2D) slices at coarse locations while keeping the high in-plane resolution in order to ensure enough body coverage while shortening the MR scan time. The aim of this study is to propose a novel method to generate HR MR images from low-resolution MR images along the longitudinal direction. In order to address the difficulty of collecting paired low- and high-resolution MR images in clinical settings and to gain the advantage of parallel cycle consistent generative adversarial networks (CycleGANs) in synthesizing realistic medical images, we developed a parallel CycleGANs based method using a self-supervised strategy.

METHODS AND MATERIALS

The proposed workflow consists of two parallely trained CycleGANs to independently predict the HR MR images in the two planes along the directions that are orthogonal to the longitudinal MR scan direction. Then, the final synthetic HR MR images are generated by fusing the two predicted images. MR images, including T1-weighted (T1), contrast enhanced T1-weighted (T1CE), T2-weighted (T2), and T2 Fluid Attenuated Inversion Recovery (FLAIR), of the multimodal brain tumor segmentation challenge 2020 (BraTS2020) dataset were processed to evaluate the proposed workflow along the cranial-caudal (CC), lateral, and anterior-posterior directions. Institutional collected MR images were also processed for evaluation of the proposed method. The performance of the proposed method was investigated via both qualitative and quantitative evaluations. Metrics of normalized mean absolute error (NMAE), peak signal-to-noise ratio (PSNR), edge keeping index (EKI), structural similarity index measurement (SSIM), information fidelity criterion (IFC), and visual information fidelity in pixel domain (VIFP) were calculated.

RESULTS

It is shown that the proposed method can generate HR MR images visually indistinguishable from the ground truth in the investigations on the BraTS2020 dataset. In addition, the intensity profiles, difference images and SSIM maps can also confirm the feasibility of the proposed method for synthesizing HR MR images. Quantitative evaluations on the BraTS2020 dataset shows that the calculated metrics of synthetic HR MR images can all be enhanced for the T1, T1CE, T2, and FLAIR images. The enhancements in the numerical metrics over the low-resolution and bi-cubic interpolated MR images, as well as those genearted with a comparative deep learning method, are statistically significant. Qualitative evaluation of the synthetic HR MR images of the clinical collected dataset could also confirm the feasibility of the proposed method.

CONCLUSIONS

The proposed method is feasible to synthesize HR MR images using self-supervised parallel CycleGANs, which can be expected to shorten MR acquisition time in clinical practices.

Bidirectional Mapping Generative Adversarial Networks for Brain MR to PET Synthesis

Fusing multi-modality medical images, such as magnetic resonance (MR) imaging and positron emission tomography (PET), can provide various anatomical and functional information about the human body. However, PET data is not always available for several reasons, such as high cost, radiation hazard, and other limitations. This paper proposes a 3D end-to-end synthesis network called Bidirectional Mapping Generative Adversarial Networks (BMGAN). Image contexts and latent vectors are effectively used for brain MR-to-PET synthesis. Specifically, a bidirectional mapping mechanism is designed to embed the semantic information of PET images into the high-dimensional latent space. Moreover, the 3D Dense-UNet generator architecture and the hybrid loss functions are further constructed to improve the visual quality of cross-modality synthetic images. The most appealing part is that the proposed method can synthesize perceptually realistic PET images while preserving the diverse brain structures of different subjects. Experimental results demonstrate that the performance of the proposed method outperforms other competitive methods in terms of quantitative measures, qualitative displays, and evaluation metrics for classification.

Data Augmentation Based on Generative Adversarial Networks to Improve Stage Classification of Chronic Kidney Disease

The prevalence of chronic kidney disease (CKD) is estimated to be 13.4% worldwide and 15% in the United States. CKD has been recognized as a leading public health problem worldwide. Unfortunately, as many as 90% of CKD patients do not know that they already have CKD. Ultrasonography is usually the first and the most commonly used imaging diagnostic tool for patients at risk of CKD. To provide a consistent assessment of the stage classifications of CKD, this study proposes an auxiliary diagnosis system based on deep learning approaches for renal ultrasound images. The system uses the ACWGAN-GP model and MobileNetV2 pre-training model. The images generated by the ACWGAN-GP generation model and the original images are simultaneously input into the pre-training model MobileNetV2 for training. This classification system achieved an accuracy of 81.9% in the four stages of CKD classification. If the prediction results allowed a higher stage tolerance, the accuracy could be improved by up to 90.1%. The proposed deep learning method solves the problem of imbalance and insufficient data samples during training processes for an automatic classification system and also improves the prediction accuracy of CKD stage diagnosis.

Generative Adversarial Networks for Morphological-Temporal Classification of Stem Cell Images

Frequently, neural network training involving biological images suffers from a lack of data, resulting in inefficient network learning. This issue stems from limitations in terms of time, resources, and difficulty in cellular experimentation and data collection. For example, when performing experimental analysis, it may be necessary for the researcher to use most of their data for testing, as opposed to model training. Therefore, the goal of this paper is to perform dataset augmentation using generative adversarial networks (GAN) to increase the classification accuracy of deep convolutional neural networks (CNN) trained on induced pluripotent stem cell microscopy images. The main challenges are: 1. modeling complex data using GAN and 2. training neural networks on augmented datasets that contain generated data. To address these challenges, a temporally constrained, hierarchical classification scheme that exploits domain knowledge is employed for model learning. First, image patches of cell colonies from gray-scale microscopy images are generated using GAN, and then these images are added to the real dataset and used to address class imbalances at multiple stages of training. Overall, a 2% increase in both true positive rate and F1-score is observed using this method as compared to a straightforward, imbalanced classification network, with some greater improvements on a classwise basis. This work demonstrates that synergistic model design involving domain knowledge is key for biological image analysis and improves model learning in high-throughput scenarios.

Feasibility of Synthetic Computed Tomography Images Generated from Magnetic Resonance Imaging Scans Using Various Deep Learning Methods in the Planning of Radiation Therapy for Prostate Cancer

Simple Summary

MRI-only simulation in radiation therapy (RT) planning has received attention because the CT scan can be omitted. For MRI-only simulation, synthetic CT (sCT) is necessary for the dose calculation. Various methodologies have been suggested for the generation of sCT and, recently, methods using the deep learning approaches are actively investigated. GAN and cycle-consistent GAN (CycGAN) have been mainly tested, however, very limited studies compared the qualities of sCTs generated from these methods or suggested other models for sCT generation. We have compared GAN, CycGAN, and, reference-guided GAN (RgGAN), a new model of deep learning method. We found that the performance in the HU conservation for soft tissue was poorest for GAN. All methods could generate sCTs feasible for VMAT planning with the trend that sCT generated from the RgGAN showed best performance in dosimetric conservation D_98% and D_95% than sCTs from other methodologies.

Abstract

We aimed to evaluate and compare the qualities of synthetic computed tomography (sCT) generated by various deep-learning methods in volumetric modulated arc therapy (VMAT) planning for prostate cancer. Simulation computed tomography (CT) and T2-weighted simulation magnetic resonance image from 113 patients were used in the sCT generation by three deep-learning approaches: generative adversarial network (GAN), cycle-consistent GAN (CycGAN), and reference-guided CycGAN (RgGAN), a new model which performed further adjustment of sCTs generated by CycGAN with available paired images. VMAT plans on the original simulation CT images were recalculated on the sCTs and the dosimetric differences were evaluated. For soft tissue, a significant difference in the mean Hounsfield unites (HUs) was observed between the original CT images and only sCTs from GAN (p = 0.03). The mean relative dose differences for planning target volumes or organs at risk were within 2% among the sCTs from the three deep-learning approaches. The differences in dosimetric parameters for D_98% and D_95% from original CT were lowest in sCT from RgGAN. In conclusion, HU conservation for soft tissue was poorest for GAN. There was the trend that sCT generated from the RgGAN showed best performance in dosimetric conservation D_98% and D_95% than sCTs from other methodologies.

Adaptive rectification based adversarial network with spectrum constraint for high-quality PET image synthesis

Positron emission tomography (PET) is a typical nuclear imaging technique, which can provide crucial functional information for early brain disease diagnosis. Generally, clinically acceptable PET images are obtained by injecting a standard-dose radioactive tracer into human body, while on the other hand the cumulative radiation exposure inevitably raises concerns about potential health risks. However, reducing the tracer dose will increase the noise and artifacts of the reconstructed PET image. For the purpose of acquiring high-quality PET images while reducing radiation exposure, in this paper, we innovatively present an adaptive rectification based generative adversarial network with spectrum constraint, named AR-GAN, which uses low-dose PET (LPET) image to synthesize standard-dose PET (SPET) image of high-quality. Specifically, considering the existing differences between the synthesized SPET image by traditional GAN and the real SPET image, an adaptive rectification network (AR-Net) is devised to estimate the residual between the preliminarily predicted image and the real SPET image, based on the hypothesis that a more realistic rectified image can be obtained by incorporating both the residual and the preliminarily predicted PET image. Moreover, to address the issue of high-frequency distortions in the output image, we employ a spectral regularization term in the training optimization objective to constrain the consistency of the synthesized image and the real image in the frequency domain, which further preserves the high-frequency detailed information and improves synthesis performance. Validations on both the phantom dataset and the clinical dataset show that the proposed AR-GAN can estimate SPET images from LPET images effectively and outperform other state-of-the-art image synthesis approaches.

Synthesis of pseudo-CT images from pelvic MRI images based on MD-CycleGAN model for radiotherapy

OBJECTIVE

A multi-discriminator-based cycle generative adversarial network (MD-CycleGAN) model was proposed to synthesize higher-quality pseudo-CT from MRI.

APPROACH

The MRI and CT images obtained at the simulation stage with cervical cancer were selected to train the model. The generator adopted the DenseNet as the main architecture. The local and global discriminators based on convolutional neural network jointly discriminated the authenticity of the input image data. In the testing phase, the model was verified by four-fold cross-validation method. In the prediction stage, the data were selected to evaluate the accuracy of the pseudo-CT in anatomy and dosimetry, and they were compared with the pseudo-CT synthesized by GAN with generator based on the architectures of ResNet, sU-Net, and FCN.

MAIN RESULTS

There are significant differences(P<0.05) in the four-fold-cross validation results on peak signal-to-noise ratio and structural similarity index metrics between the pseudo-CT obtained based on MD-CycleGAN and the ground truth CT (CTgt). The pseudo-CT synthesized by MD-CycleGAN had closer anatomical information to the CTgt with root mean square error of 47.83±2.92 HU and normalized mutual information value of 0.9014±0.0212 and mean absolute error value of 46.79±2.76 HU. The differences in dose distribution between the pseudo-CT obtained by MD-CycleGAN and the CTgt were minimal. The mean absolute dose errors of Dosemax, Dosemin and Dosemean based on the planning target volume were used to evaluate the dose uncertainty of the four pseudo-CT. The u-values of the Wilcoxon test were 55.407, 41.82 and 56.208, and the differences were statistically significant. The 2%/2 mm-based gamma pass rate (%) of the proposed method was 95.45±1.91, and the comparison methods (ResNet_GAN, sUnet_GAN and FCN_GAN) were 93.33±1.20, 89.64±1.63 and 87.31±1.94, respectively.

SIGNIFICANCE

The pseudo-CT obtained based on MD-CycleGAN have higher imaging quality and are closer to the CTgt in terms of anatomy and dosimetry than other GAN models.

Deep learning based MRI contrast synthesis using full volume prediction using full volume prediction

In Magnetic Resonance Imaging (MRI), depending on the image acquisition settings, a large number of image types or contrasts can be generated showing complementary information of the same imaged subject. This multi-spectral information is highly beneficial since can improve MRI analysis tasks such as segmentation and registration, thanks to pattern ambiguity reduction. However, the acquisition of several contrasts is not always possible due to time limitations and patient comfort constraints. Contrast synthesis has emerged recently as an approximate solution to generate other image types different from those acquired originally. Most of the previously proposed methods for contrast synthesis are slice-based which result in intensity inconsistencies between neighbor slices when applied in 3D. We propose the use of a 3D convolutional neural network (CNN) capable of generating T2 and FLAIR images from a single anatomical T1 source volume. The proposed network is a 3D variant of the UNet that processes the whole volume at once breaking with the inconsistency in the resulting output volumes related to 2D slice or patch-based methods. Since working with a full volume at once has a huge memory demand we have introduced a spatial-to-depth and a reconstruction layer that allows working with the full volume but maintain the required network complexity to solve the problem. Our approach enhances the coherence in the synthesized volume while improving the accuracy thanks to the integrated three-dimensional context-awareness. Finally, the proposed method has been validated with a segmentation method, thus demonstrating its usefulness in a direct and relevant application.

Advancing diagnostic performance and clinical applicability of deep learning-driven generative adversarial networks for Alzheimer's disease

Alzheimer's disease (AD) is a neurodegenerative disease that severely affects the activities of daily living in aged individuals, which typically needs to be diagnosed at an early stage. Generative adversarial networks (GANs) provide a new deep learning method that show good performance in image processing, while it remains to be verified whether a GAN brings benefit in AD diagnosis. The purpose of this research is to systematically review psychoradiological studies on the application of a GAN in the diagnosis of AD from the aspects of classification of AD state and AD-related image processing compared with other methods. In addition, we evaluated the research methodology and provided suggestions from the perspective of clinical application. Compared with other methods, a GAN has higher accuracy in the classification of AD state and better performance in AD-related image processing (e.g. image denoising and segmentation). Most studies used data from public databases but lacked clinical validation, and the process of quantitative assessment and comparison in these studies lacked clinicians' participation, which may have an impact on the improvement of generation effect and generalization ability of the GAN model. The application value of GANs in the classification of AD state and AD-related image processing has been confirmed in reviewed studies. Improvement methods toward better GAN architecture were also discussed in this paper. In sum, the present study demonstrated advancing diagnostic performance and clinical applicability of GAN for AD, and suggested that the future researchers should consider recruiting clinicians to compare the algorithm with clinician manual methods and evaluate the clinical effect of the algorithm.

Automatic Inter-Frame Patient Motion Correction for Dynamic Cardiac PET Using Deep Learning

Patient motion during dynamic PET imaging can induce errors in myocardial blood flow (MBF) estimation. Motion correction for dynamic cardiac PET is challenging because the rapid tracer kinetics of 82Rb leads to substantial tracer distribution change across different dynamic frames over time, which can cause difficulties for image registration-based motion correction, particularly for early dynamic frames. In this paper, we developed an automatic deep learning-based motion correction (DeepMC) method for dynamic cardiac PET. In this study we focused on the detection and correction of inter-frame rigid translational motion caused by voluntary body movement and pattern change of respiratory motion. A bidirectional-3D LSTM network was developed to fully utilize both local and nonlocal temporal information in the 4D dynamic image data for motion detection. The network was trained and evaluated over motion-free patient scans with simulated motion so that the motion ground-truths are available, where one million samples based on 65 patient scans were used in training, and 600 samples based on 20 patient scans were used in evaluation. The proposed method was also evaluated using additional 10 patient datasets with real motion. We demonstrated that the proposed DeepMC obtained superior performance compared to conventional registration-based methods and other convolutional neural networks (CNN), in terms of motion estimation and MBF quantification accuracy. Once trained, DeepMC is much faster than the registration-based methods and can be easily integrated into the clinical workflow. In the future work, additional investigation is needed to evaluate this approach in a clinical context with realistic patient motion.

Improving Skin Cancer Classification Using Heavy-Tailed Student T-Distribution in Generative Adversarial Networks (TED-GAN)

Deep learning has gained immense attention from researchers in medicine, especially in medical imaging. The main bottleneck is the unavailability of sufficiently large medical datasets required for the good performance of deep learning models. This paper proposes a new framework consisting of one variational autoencoder (VAE), two generative adversarial networks, and one auxiliary classifier to artificially generate realistic-looking skin lesion images and improve classification performance. We first train the encoder-decoder network to obtain the latent noise vector with the image manifold’s information and let the generative adversarial network sample the input from this informative noise vector in order to generate the skin lesion images. The use of informative noise allows the GAN to avoid mode collapse and creates faster convergence. To improve the diversity in the generated images, we use another GAN with an auxiliary classifier, which samples the noise vector from a heavy-tailed student t-distribution instead of a random noise Gaussian distribution. The proposed framework was named TED-GAN, with T from the t-distribution and ED from the encoder-decoder network which is part of the solution. The proposed framework could be used in a broad range of areas in medical imaging. We used it here to generate skin lesion images and have obtained an improved classification performance on the skin lesion classification task, rising from 66% average accuracy to 92.5%. The results show that TED-GAN has a better impact on the classification task because of its diverse range of generated images due to the use of a heavy-tailed t-distribution.

Deep Learning for the Industrial Internet of Things (IIoT): A Comprehensive Survey of Techniques, Implementation Frameworks, Potential Applications, and Future Directions

The Industrial Internet of Things (IIoT) refers to the use of smart sensors, actuators, fast communication protocols, and efficient cybersecurity mechanisms to improve industrial processes and applications. In large industrial networks, smart devices generate large amounts of data, and thus IIoT frameworks require intelligent, robust techniques for big data analysis. Artificial intelligence (AI) and deep learning (DL) techniques produce promising results in IIoT networks due to their intelligent learning and processing capabilities. This survey article assesses the potential of DL in IIoT applications and presents a brief architecture of IIoT with key enabling technologies. Several well-known DL algorithms are then discussed along with their theoretical backgrounds and several software and hardware frameworks for DL implementations. Potential deployments of DL techniques in IIoT applications are briefly discussed. Finally, this survey highlights significant challenges and future directions for future research endeavors.

Synthetic digital reconstructed radiographs for MR-only robotic stereotactic radiation therapy: A proof of concept

PURPOSE

To create synthetic CTs and digital reconstructed radiographs (DRRs) from MR images that allow for fiducial visualization and accurate dose calculation for MR-only radiosurgery.

METHODS

We developed a machine learning model to create synthetic CTs from pelvic MRs for prostate treatments. This model has been previously proven to generate synthetic CTs with accuracy on par or better than alternate methods, such as atlas-based registration. Our dataset consisted of 11 paired CT and conventional MR (T2) images used for previous CyberKnife (Accuray, Inc) radiotherapy treatments. The MR images were pre-processed to mimic the appearance of fiducial-enhancing images. Two models were trained for each parameter case, using a sub-set of the available image pairs, with the remaining images set aside for testing and validation of the model to identify the optimal patch size and number of image pairs used for training. Four models were then trained using the identified parameters and used to generate synthetic CTs, which in turn were used to generate DRRs at angles 45° and 315°, as would be used for a CyberKnife treatment. The synthetic CTs and DRRs were compared visually and using the mean squared error and peak signal-to-noise ratio against the ground-truth images to evaluate their similarity.

RESULTS

The synthetic CTs, as well as the DRRs generated from them, gave similar visualization of the fiducial markers in the prostate as the true counterparts. There was no significant difference found for the fiducial localization for the CTs and DRRs. Across the 8 DRRs analyzed, the mean MSE between the normalized true and synthetic DRRs was 0.66 ± 0.42% and the mean PSNR for this region was 22.9 ± 3.7 dB. For the full CTs, the mean MAE was 72.9 ± 88.1 HU and the mean PSNR was 31.2 ± 2.2 dB.

CONCLUSIONS

Our machine learning-based method provides a proof of concept of a way to generate synthetic CTs and DRRs for accurate dose calculation and fiducial localization for use in radiation treatment of the prostate.

Creating High-Resolution Microscopic Cross-Section Images of Hardwood Species Using Generative Adversarial Networks

Microscopic wood identification plays a critical role in many economically important areas in wood science. Historically, producing and curating relevant and representative microscopic cross-section images of wood species is limited to highly experienced and trained anatomists. This manuscript demonstrates the feasibility of generating synthetic microscopic cross-sections of hardwood species. We leveraged a publicly available dataset of 119 hardwood species to train a style-based generative adversarial network (GAN). The proposed GAN generated anatomically accurate cross-section images with remarkable fidelity to actual data. Quantitative metrics corroborated the capacity of the generative model in capturing complex wood structure by resulting in a Fréchet inception distance score of 17.38. Image diversity was calculated using the Structural Similarity Index Measure (SSIM). The SSIM results confirmed that the GAN approach can successfully synthesize diverse images. To confirm the usefulness and realism of the GAN generated images, eight professional wood anatomists in two experience levels participated in a visual Turing test and correctly identified fake and actual images at rates of 48.3 and 43.7%, respectively, with no statistical difference when compared to random guess. The generative model can synthesize realistic, diverse, and meaningful high-resolution microscope cross-section images that are virtually indistinguishable from real images. Furthermore, the framework presented may be suitable for improving current deep learning models, helping understand potential breeding between species, and may be used as an educational tool.

Comparison of different deep learning architectures for synthetic CT generation from MR images

PURPOSE

Among the different available methods for synthetic CT generation from MR images for the task of MR-guided radiation planning, the deep learning algorithms have and do outperform their conventional counterparts. In this study, we investigated the performance of some most popular deep learning architectures including eCNN, U-Net, GAN, V-Net, and Res-Net for the task of sCT generation. As a baseline, an atlas-based method is implemented to which the results of the deep learning-based model are compared.

METHODS

A dataset consisting of 20 co-registered MR-CT pairs of the male pelvis is applied to assess the different sCT production methods' performance. The mean error (ME), mean absolute error (MAE), Pearson correlation coefficient (PCC), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) metrics were computed between the estimated sCT and the ground truth (reference) CT images.

RESULTS

The visual inspection revealed that the sCTs produced by eCNN, V-Net, and ResNet, unlike the other methods, were less noisy and greatly resemble the ground truth CT image. In the whole pelvis region, the eCNN yielded the lowest MAE (26.03 ± 8.85 HU) and ME (0.82 ± 7.06 HU), and the highest PCC metrics were yielded by the eCNN (0.93 ± 0.05) and ResNet (0.91 ± 0.02) methods. The ResNet model had the highest PSNR of 29.38 ± 1.75 among all models. In terms of the Dice similarity coefficient, the eCNN method revealed superior performance in major tissue identification (air, bone, and soft tissue).

CONCLUSIONS

All in all, the eCNN and ResNet deep learning methods revealed acceptable performance with clinically tolerable quantification errors.

Bidirectional Mapping-Based Domain Adaptation for Nucleus Detection in Cross-Modality Microscopy Images

Cell or nucleus detection is a fundamental task in microscopy image analysis and has recently achieved state-of-the-art performance by using deep neural networks. However, training supervised deep models such as convolutional neural networks (CNNs) usually requires sufficient annotated image data, which is prohibitively expensive or unavailable in some applications. Additionally, when applying a CNN to new datasets, it is common to annotate individual cells/nuclei in those target datasets for model re-learning, leading to inefficient and low-throughput image analysis. To tackle these problems, we present a bidirectional, adversarial domain adaptation method for nucleus detection on cross-modality microscopy image data. Specifically, the method learns a deep regression model for individual nucleus detection with both source-to-target and target-to-source image translation. In addition, we explicitly extend this unsupervised domain adaptation method to a semi-supervised learning situation and further boost the nucleus detection performance. We evaluate the proposed method on three cross-modality microscopy image datasets, which cover a wide variety of microscopy imaging protocols or modalities, and obtain a significant improvement in nucleus detection compared to reference baseline approaches. In addition, our semi-supervised method is very competitive with recent fully supervised learning models trained with all real target training labels.

Active Cell Appearance Model Induced Generative Adversarial Networks for Annotation-Efficient Cell Segmentation and Identification on Adaptive Optics Retinal Images

Data annotation is a fundamental precursor for establishing large training sets to effectively apply deep learning methods to medical image analysis. For cell segmentation, obtaining high quality annotations is an expensive process that usually requires manual grading by experts. This work introduces an approach to efficiently generate annotated images, called "A-GANs", created by combining an active cell appearance model (ACAM) with conditional generative adversarial networks (C-GANs). ACAM is a statistical model that captures a realistic range of cell characteristics and is used to ensure that the image statistics of generated cells are guided by real data. C-GANs utilize cell contours generated by ACAM to produce cells that match input contours. By pairing ACAM-generated contours with A-GANs-based generated images, high quality annotated images can be efficiently generated. Experimental results on adaptive optics (AO) retinal images showed that A-GANs robustly synthesize realistic, artificial images whose cell distributions are exquisitely specified by ACAM. The cell segmentation performance using as few as 64 manually-annotated real AO images combined with 248 artificially-generated images from A-GANs was similar to the case of using 248 manually-annotated real images alone (Dice coefficients of 88% for both). Finally, application to rare diseases in which images exhibit never-seen characteristics demonstrated improvements in cell segmentation without the need for incorporating manual annotations from these new retinal images. Overall, A-GANs introduce a methodology for generating high quality annotated data that statistically captures the characteristics of any desired dataset and can be used to more efficiently train deep-learning-based medical image analysis applications.

Pathological neural networks and artificial neural networks in ALS: diagnostic classification based on pathognomonic neuroimaging features

The description of group-level, genotype- and phenotype-associated imaging traits is academically important, but the practical demands of clinical neurology centre on the accurate classification of individual patients into clinically relevant diagnostic, prognostic and phenotypic categories. Similarly, pharmaceutical trials require the precision stratification of participants based on quantitative measures. A single-centre study was conducted with a uniform imaging protocol to test the accuracy of an artificial neural network classification scheme on a cohort of 378 participants composed of patients with ALS, healthy subjects and disease controls. A comprehensive panel of cerebral volumetric measures, cortical indices and white matter integrity values were systematically retrieved from each participant and fed into a multilayer perceptron model. Data were partitioned into training and testing and receiver-operating characteristic curves were generated for the three study-groups. Area under the curve values were 0.930 for patients with ALS, 0.958 for disease controls, and 0.931 for healthy controls relying on all input imaging variables. The ranking of variables by classification importance revealed that white matter metrics were far more relevant than grey matter indices to classify single subjects. The model was further tested in a subset of patients scanned within 6 weeks of their diagnosis and an AUC of 0.915 was achieved. Our study indicates that individual subjects may be accurately categorised into diagnostic groups in an observer-independent classification framework based on multiparametric, spatially registered radiology data. The development and validation of viable computational models to interpret single imaging datasets are urgently required for a variety of clinical and clinical trial applications.

Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation

Privacy concerns around sharing personally identifiable information are a major barrier to data sharing in medical research. In many cases, researchers have no interest in a particular individual's information but rather aim to derive insights at the level of cohorts. Here, we utilise generative adversarial networks (GANs) to create medical imaging datasets consisting entirely of synthetic patient data. The synthetic images ideally have, in aggregate, similar statistical properties to those of a source dataset but do not contain sensitive personal information. We assess the quality of synthetic data generated by two GAN models for chest radiographs with 14 radiology findings and brain computed tomography (CT) scans with six types of intracranial haemorrhages. We measure the synthetic image quality by the performance difference of predictive models trained on either the synthetic or the real dataset. We find that synthetic data performance disproportionately benefits from a reduced number of classes. Our benchmark also indicates that at low numbers of samples per class, label overfitting effects start to dominate GAN training. We conducted a reader study in which trained radiologists discriminate between synthetic and real images. In accordance with our benchmark results, the classification accuracy of radiologists improves with an increasing resolution. Our study offers valuable guidelines and outlines practical conditions under which insights derived from synthetic images are similar to those that would have been derived from real data. Our results indicate that synthetic data sharing may be an attractive alternative to sharing real patient-level data in the right setting.

Realistic image normalization for multi-Domain segmentation

Image normalization is a building block in medical image analysis. Conventional approaches are customarily employed on a per-dataset basis. This strategy, however, prevents the current normalization algorithms from fully exploiting the complex joint information available across multiple datasets. Consequently, ignoring such joint information has a direct impact on the processing of segmentation algorithms. This paper proposes to revisit the conventional image normalization approach by, instead, learning a common normalizing function across multiple datasets. Jointly normalizing multiple datasets is shown to yield consistent normalized images as well as an improved image segmentation when intensity shifts are large. To do so, a fully automated adversarial and task-driven normalization approach is employed as it facilitates the training of realistic and interpretable images while keeping performance on par with the state-of-the-art. The adversarial training of our network aims at finding the optimal transfer function to improve both, jointly, the segmentation accuracy and the generation of realistic images. We have evaluated the performance of our normalizer on both infant and adult brain images from the iSEG, MRBrainS and ABIDE datasets. The results indicate that our contribution does provide an improved realism to the normalized images, while retaining a segmentation accuracy at par with the state-of-the-art learnable normalization approaches.

Realistic Lung Nodule Synthesis With Multi-Target Co-Guided Adversarial Mechanism

The important cues for a realistic lung nodule synthesis include the diversity in shape and background, controllability of semantic feature levels, and overall CT image quality. To incorporate these cues as the multiple learning targets, we introduce the Multi-Target Co-Guided Adversarial Mechanism, which utilizes the foreground and background mask to guide nodule shape and lung tissues, takes advantage of the CT lung and mediastinal window as the guidance of spiculation and texture control, respectively. Further, we propose a Multi-Target Co-Guided Synthesizing Network with a joint loss function to realize the co-guidance of image generation and semantic feature learning. The proposed network contains a Mask-Guided Generative Adversarial Sub-Network (MGGAN) and a Window-Guided Semantic Learning Sub-Network (WGSLN). The MGGAN generates the initial synthesis using the mask combined with the foreground and background masks, guiding the generation of nodule shape and background tissues. Meanwhile, the WGSLN controls the semantic features and refines the synthesis quality by transforming the initial synthesis into the CT lung and mediastinal window, and performing the spiculation and texture learning simultaneously. We validated our method using the quantitative analysis of authenticity under the Fréchet Inception Score, and the results show its state-of-the-art performance. We also evaluated our method as a data augmentation method to predict malignancy level on the LIDC-IDRI database, and the results show that the accuracy of VGG-16 is improved by 5.6%. The experimental results confirm the effectiveness of the proposed method.

Generation of synthetic PET images of synaptic density and amyloid from 18 F-FDG images using deep learning

PURPOSE

Positron emission tomography (PET) imaging with various tracers is increasingly used in Alzheimer's disease (AD) studies. However, access to PET scans using new or less-available tracers with sophisticated synthesis and short half-life isotopes may be very limited. Therefore, it is of great significance and interest in AD research to assess the feasibility of generating synthetic PET images of less-available tracers from the PET image of another common tracer, in particular 18 F-FDG.

METHODS

We implemented advanced deep learning methods using the U-Net model to predict 11 C-UCB-J PET images of synaptic vesicle protein 2A (SV2A), a surrogate of synaptic density, from 18 F-FDG PET data. Dynamic 18 F-FDG and 11 C-UCB-J scans were performed in 21 participants with normal cognition (CN) and 33 participants with Alzheimer's disease (AD). Cerebellum was used as the reference region for both tracers. For 11 C-UCB-J image prediction, four network models were trained and tested, which included 1) 18 F-FDG SUV ratio (SUVR) to 11 C-UCB-J SUVR, 2) 18 F-FDG Ki ratio to 11 C-UCB-J SUVR, 3) 18 F-FDG SUVR to 11 C-UCB-J distribution volume ratio (DVR), and 4) 18 F-FDG Ki ratio to 11 C-UCB-J DVR. The normalized root mean square error (NRMSE), structure similarity index (SSIM), and Pearson's correlation coefficient were calculated for evaluating the overall image prediction accuracy. Mean bias of various ROIs in the brain and correlation plots between predicted images and true images were calculated for ROI-based prediction accuracy. Following a similar training and evaluation strategy, 18 F-FDG SUVR to 11 C-PiB SUVR network was also trained and tested for 11 C-PiB static image prediction.

RESULTS

The results showed that all four network models obtained satisfactory 11 C-UCB-J static and parametric images. For 11 C-UCB-J SUVR prediction, the mean ROI bias was -0.3% ± 7.4% for the AD group and -0.5% ± 7.3% for the CN group with 18 F-FDG SUVR as the input, -0.7% ± 8.1% for the AD group, and -1.3% ± 7.0% for the CN group with 18 F-FDG Ki ratio as the input. For 11 C-UCB-J DVR prediction, the mean ROI bias was -1.3% ± 7.5% for the AD group and -2.0% ± 6.9% for the CN group with 18 F-FDG SUVR as the input, -0.7% ± 9.0% for the AD group, and -1.7% ± 7.8% for the CN group with 18 F-FDG Ki ratio as the input. For 11 C-PiB SUVR image prediction, which appears to be a more challenging task, the incorporation of additional diagnostic information into the network is needed to control the bias below 5% for most ROIs.

CONCLUSIONS

It is feasible to use 3D U-Net-based methods to generate synthetic 11 C-UCB-J PET images from 18 F-FDG images with reasonable prediction accuracy. It is also possible to predict 11 C-PiB SUVR images from 18 F-FDG images, though the incorporation of additional non-imaging information is needed.

Normalization of breast MRIs using cycle-consistent generative adversarial networks

OBJECTIVES

Dynamic Contrast Enhanced-Magnetic Resonance Imaging (DCE-MRI) is widely used to complement ultrasound examinations and x-ray mammography for early detection and diagnosis of breast cancer. However, images generated by various MRI scanners (e.g., GE Healthcare, and Siemens) differ both in intensity and noise distribution, preventing algorithms trained on MRIs from one scanner to generalize to data from other scanners. In this work, we propose a method to solve this problem by normalizing images between various scanners.

METHODS

MRI normalization is challenging because it requires normalizing intensity values and mapping noise distributions between scanners. We utilize a cycle-consistent generative adversarial network to learn a bidirectional mapping and perform normalization between MRIs produced by GE Healthcare and Siemens scanners in an unpaired setting. Initial experiments demonstrate that the traditional CycleGAN architecture struggles to preserve the anatomical structures of the breast during normalization. Thus, we propose two technical innovations in order to preserve both the shape of the breast as well as the tissue structures within the breast. First, we incorporate mutual information loss during training in order to ensure anatomical consistency. Second, we propose a modified discriminator architecture that utilizes a smaller field-of-view to ensure the preservation of finer details in the breast tissue.

RESULTS

Quantitative and qualitative evaluations show that the second innovation consistently preserves the breast shape and tissue structures while also performing the proper intensity normalization and noise distribution mapping.

CONCLUSION

Our results demonstrate that the proposed model can successfully learn a bidirectional mapping and perform normalization between MRIs produced by different vendors, potentially enabling improved diagnosis and detection of breast cancer. All the data used in this study are publicly available at https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70226903.

Paired-unpaired Unsupervised Attention Guided GAN with transfer learning for bidirectional brain MR-CT synthesis

Medical image acquisition plays a significant role in the diagnosis and management of diseases. Magnetic Resonance (MR) and Computed Tomography (CT) are considered two of the most popular modalities for medical image acquisition. Some considerations, such as cost and radiation dose, may limit the acquisition of certain image modalities. Therefore, medical image synthesis can be used to generate required medical images without actual acquisition. In this paper, we propose a paired-unpaired Unsupervised Attention Guided Generative Adversarial Network (uagGAN) model to translate MR images to CT images and vice versa. The uagGAN model is pre-trained with a paired dataset for initialization and then retrained on an unpaired dataset using a cascading process. In the paired pre-training stage, we enhance the loss function of our model by combining the Wasserstein GAN adversarial loss function with a new combination of non-adversarial losses (content loss and L1) to generate fine structure images. This will ensure global consistency, and better capture of the high and low frequency details of the generated images. The uagGAN model is employed as it generates more accurate and sharper images through the production of attention masks. Knowledge from a non-medical pre-trained model is also transferred to the uagGAN model for improved learning and better image translation performance. Quantitative evaluation and qualitative perceptual analysis by radiologists indicate that employing transfer learning with the proposed paired-unpaired uagGAN model can achieve better performance as compared to other rival image-to-image translation models.

Improvement of megavoltage computed tomography image quality for adaptive helical tomotherapy using cycleGAN-based image synthesis with small datasets

PURPOSE

Megavoltage computed tomography (MVCT) offers an opportunity for adaptive helical tomotherapy. However, high noise and reduced contrast in the MVCT images due to a decrease in the imaging dose to patients limits its usability. Therefore, we propose an algorithm to improve the image quality of MVCT.

METHODS

The proposed algorithm generates kilovoltage CT (kVCT)-like images from MVCT images using a cycle-consistency generative adversarial network (cycleGAN)-based image synthesis model. Data augmentation using an affine transformation was applied to the training data to overcome the lack of data diversity in the network training. The mean absolute error (MAE), root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) were used to quantify the correction accuracy of the images generated by the proposed algorithm. The proposed method was validated by comparing the images generated with those obtained from conventional and deep learning-based image processing method through non-augmented datasets.

RESULTS

The average MAE, RMSE, PSNR, and SSIM values were 18.91 HU, 69.35 HU, 32.73 dB, and 95.48 using the proposed method, respectively, whereas cycleGAN with non-augmented data showed inferior results (19.88 HU, 70.55 HU, 32.62 dB, 95.19, respectively). The voxel values of the image obtained by the proposed method also indicated similar distributions to those of the kVCT image. The dose-volume histogram of the proposed method was also similar to that of electron density corrected MVCT.

CONCLUSIONS

The proposed algorithm generates synthetic kVCT images from MVCT images using cycleGAN with small patient datasets. The image quality achieved by the proposed method was correspondingly improved to the level of a kVCT image while maintaining the anatomical structure of an MVCT image. The evaluation of dosimetric effectiveness of the proposed method indicates the applicability of accurate treatment planning in adaptive radiation therapy.

Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast

Most existing algorithms for automatic 3D morphometry of human brain MRI scans are designed for data with near-isotropic voxels at approximately 1 mm resolution, and frequently have contrast constraints as well-typically requiring T1-weighted images (e.g., MP-RAGE scans). This limitation prevents the analysis of millions of MRI scans acquired with large inter-slice spacing in clinical settings every year. In turn, the inability to quantitatively analyze these scans hinders the adoption of quantitative neuro imaging in healthcare, and also precludes research studies that could attain huge sample sizes and hence greatly improve our understanding of the human brain. Recent advances in convolutional neural networks (CNNs) are producing outstanding results in super-resolution and contrast synthesis of MRI. However, these approaches are very sensitive to the specific combination of contrast, resolution and orientation of the input images, and thus do not generalize to diverse clinical acquisition protocols - even within sites. In this article, we present SynthSR, a method to train a CNN that receives one or more scans with spaced slices, acquired with different contrast, resolution and orientation, and produces an isotropic scan of canonical contrast (typically a 1 mm MP-RAGE). The presented method does not require any preprocessing, beyond rigid coregistration of the input scans. Crucially, SynthSR trains on synthetic input images generated from 3D segmentations, and can thus be used to train CNNs for any combination of contrasts, resolutions and orientations without high-resolution real images of the input contrasts. We test the images generated with SynthSR in an array of common downstream analyses, and show that they can be reliably used for subcortical segmentation and volumetry, image registration (e.g., for tensor-based morphometry), and, if some image quality requirements are met, even cortical thickness morphometry. The source code is publicly available at https://github.com/BBillot/SynthSR.

Acoustic simulation for transcranial focused ultrasound using GAN-based synthetic CT

Transcranial focused ultrasound (tFUS) is a promising non-invasive technique for treating neurological and psychiatric disorders. One of the challenges for tFUS is the disruption of wave propagation through the skull. Consequently, despite the risks associated with exposure to ionizing radiation, computed tomography (CT) is required to estimate the acoustic transmission through the skull. This study aims to generate synthetic CT (sCT) from T1-weighted magnetic resonance imaging (MRI) and investigate its applicability to tFUS acoustic simulation. We trained a 3D conditional generative adversarial network (3D-cGAN) with 15 subjects. We then assessed image quality with 15 test subjects: mean absolute error (MAE) = 85.72± 9.50 HU (head) and 280.25±24.02 HU (skull), dice coefficient similarity (DSC) = 0.88±0.02 (skull). In terms of skull density ratio (SDR) and skull thickness (ST), no significant difference was found between sCT and real CT (rCT). When the acoustic simulation results of rCT and sCT were compared, the intracranial peak acoustic pressure ratio was found to be less than 4%, and the distance between focal points less than 1 mm.

Fat-saturated image generation from multi-contrast MRIs using generative adversarial networks with Bloch equation-based autoencoder regularization

Obtaining multiple series of magnetic resonance (MR) images with different contrasts is useful for accurate diagnosis of human spinal conditions. However, this can be time consuming and a burden on both the patient and the hospital. We propose a Bloch equation-based autoencoder regularization generative adversarial network (BlochGAN) to generate a fat saturation T2-weighted (T2 FS) image from T1-weighted (T1-w) and T2-weighted (T2-w) images of human spine. To achieve this, our approach was to utilize the relationship between the contrasts using Bloch equation since it is a fundamental principle of MR physics and serves as a physical basis of each contrasts. BlochGAN properly generated the target-contrast images using the autoencoder regularization based on the Bloch equation to identify the physical basis of the contrasts. BlochGAN consists of four sub-networks: an encoder, a decoder, a generator, and a discriminator. The encoder extracts features from the multi-contrast input images, and the generator creates target T2 FS images using the features extracted from the encoder. The discriminator assists network learning by providing adversarial loss, and the decoder reconstructs the input multi-contrast images and regularizes the learning process by providing reconstruction loss. The discriminator and the decoder are only used in the training process. Our results demonstrate that BlochGAN achieved quantitatively and qualitatively superior performance compared to conventional medical image synthesis methods in generating spine T2 FS images from T1-w, and T2-w images.

High through-plane resolution CT imaging with self-supervised deep learning

CT images for radiotherapy planning are usually acquired in thick slices to reduce the imaging dose, especially for pediatric patients, and to lessen the need for contouring and treatment planning on more slices. However, low through-plane resolution may degrade the accuracy of dose calculations. In this paper, a self-supervised deep learning workflow is proposed to synthesize high through-plane resolution CT images by learning from their high in-plane resolution features. The proposed workflow was designed to facilitate neural networks to learn the mapping from low-resolution (LR) to high-resolution (HR) images in the axial plane. During the inference step, the HR sagittal and coronal images were generated by feeding two parallel-trained neural networks with the respective LR sagittal and coronal images to the trained neural networks. The CT simulation images of a cohort of 75 patients with head and neck cancer (1 mm slice thickness) and 200 CT images of a cohort of 20 lung cancer patients (3 mm slice thickness) were retrospectively investigated in a cross-validation manner. The HR images generated with the proposed method were qualitatively (visual quality, image intensity profiles and preliminary observer study) and quantitatively (mean absolute error, edge keeping index, structural similarity index measurement, information fidelity criterion and visual information fidelity in pixel domain) inspected, while taking the original CT images of the head and neck and lung cancer patients as the reference. The qualitative results showed the capability of the proposed method for generating high through-plane resolution CT images with data from both groups of cancer patients. All the improvements in the measure metrics were confirmed to be statistically significant with paired two-samplet-test analysis. The innovative point of the work is that the proposed deep learning workflow for CT image generation with high through-plane resolution in radiotherapy is self-supervised, meaning that it does not rely on ground truth CT images to train the network. In addition, the assumption that the in-plane HR information can supervise the through-plane HR generation is confirmed. We hope that this will inspire more research on this topic to further improve the through-plane resolution of medical images.

Deep learning-based T1-enhanced selection of linear attenuation coefficients (DL-TESLA) for PET/MR attenuation correction in dementia neuroimaging

PURPOSE

The accuracy of existing PET/MR attenuation correction (AC) has been limited by a lack of correlation between MR signal and tissue electron density. Based on our finding that longitudinal relaxation rate, or R1 , is associated with CT Hounsfield unit in bone and soft tissues in the brain, we propose a deep learning T1 -enhanced selection of linear attenuation coefficients (DL-TESLA) method to incorporate quantitative R1 for PET/MR AC and evaluate its accuracy and longitudinal test-retest repeatability in brain PET/MR imaging.

METHODS

DL-TESLA uses a 3D residual UNet (ResUNet) for pseudo-CT (pCT) estimation. With a total of 174 participants, we compared PET AC accuracy of DL-TESLA to 3 other methods adopting similar 3D ResUNet structures but using UTE R 2 ∗ , or Dixon, or T1 -MPRAGE as input. With images from 23 additional participants repeatedly scanned, the test-retest differences and within-subject coefficient of variation of standardized uptake value ratios (SUVR) were compared between PET images reconstructed using either DL-TESLA or CT for AC.

RESULTS

DL-TESLA had (1) significantly lower mean absolute error in pCT, (2) the highest Dice coefficients in both bone and air, (3) significantly lower PET relative absolute error in whole brain and various brain regions, (4) the highest percentage of voxels with a PET relative error within both ±3% and ±5%, (5) similar to CT test-retest differences in SUVRs from the cerebrum and mean cortical (MC) region, and (6) similar to CT within-subject coefficient of variation in cerebrum and MC.

CONCLUSION

DL-TESLA demonstrates excellent PET/MR AC accuracy and test-retest repeatability.

GANDA: A deep generative adversarial network conditionally generates intratumoral nanoparticles distribution pixels-to-pixels

Intratumoral nanoparticles (NPs) distribution is critical for the success of nanomedicine in imaging and treatment, but computational models to describe the NPs distribution remain unavailable due to the complex tumor-nano interactions. Here, we develop a Generative Adversarial Network for Distribution Analysis (GANDA) to describe and conditionally generates the intratumoral quantum dots (QDs) distribution after i.v. injection. This deep generative model is trained automatically by 27,775 patches of tumor vessels and cell nuclei decomposed from whole-slide images of 4 T1 breast cancer sections. The GANDA model can conditionally generate images of intratumoral QDs distribution under the constraint of given tumor vessels and cell nuclei channels with the same spatial resolution (pixels-to-pixels), minimal loss (mean squared error, MSE = 1.871) and excellent reliability (intraclass correlation, ICC = 0.94). Quantitative analysis of QDs extravasation distance (ICC = 0.95) and subarea distribution (ICC = 0.99) is allowed on the generated images without knowing the real QDs distribution. We believe this deep generative model may provide opportunities to investigate how influencing factors affect NPs distribution in individual tumors and guide nanomedicine optimization for molecular imaging and personalized treatment.

Recent Radiomics Advancements in Breast Cancer: Lessons and Pitfalls for the Next Future

Radiomics is an emerging translational field of medicine based on the extraction of high-dimensional data from radiological images, with the purpose to reach reliable models to be applied into clinical practice for the purposes of diagnosis, prognosis and evaluation of disease response to treatment. We aim to provide the basic information on radiomics to radiologists and clinicians who are focused on breast cancer care, encouraging cooperation with scientists to mine data for a better application in clinical practice. We investigate the workflow and clinical application of radiomics in breast cancer care, as well as the outlook and challenges based on recent studies. Currently, radiomics has the potential ability to distinguish between benign and malignant breast lesions, to predict breast cancer’s molecular subtypes, the response to neoadjuvant chemotherapy and the lymph node metastases. Even though radiomics has been used in tumor diagnosis and prognosis, it is still in the research phase and some challenges need to be faced to obtain a clinical translation. In this review, we discuss the current limitations and promises of radiomics for improvement in further research.

MR-contrast-aware image-to-image translations with generative adversarial networks

Purpose

A magnetic resonance imaging (MRI) exam typically consists of several sequences that yield different image contrasts. Each sequence is parameterized through multiple acquisition parameters that influence image contrast, signal-to-noise ratio, acquisition time, and/or resolution. Depending on the clinical indication, different contrasts are required by the radiologist to make a diagnosis. As MR sequence acquisition is time consuming and acquired images may be corrupted due to motion, a method to synthesize MR images with adjustable contrast properties is required.

Methods

Therefore, we trained an image-to-image generative adversarial network conditioned on the MR acquisition parameters repetition time and echo time. Our approach is motivated by style transfer networks, whereas the “style” for an image is explicitly given in our case, as it is determined by the MR acquisition parameters our network is conditioned on.

Results

This enables us to synthesize MR images with adjustable image contrast. We evaluated our approach on the fastMRI dataset, a large set of publicly available MR knee images, and show that our method outperforms a benchmark pix2pix approach in the translation of non-fat-saturated MR images to fat-saturated images. Our approach yields a peak signal-to-noise ratio and structural similarity of 24.48 and 0.66, surpassing the pix2pix benchmark model significantly.

Conclusion

Our model is the first that enables fine-tuned contrast synthesis, which can be used to synthesize missing MR-contrasts or as a data augmentation technique for AI training in MRI. It can also be used as basis for other image-to-image translation tasks within medical imaging, e.g., to enhance intermodality translation (MRI → CT) or 7 T image synthesis from 3 T MR images.

Multi-modal MRI Image Synthesis via GAN with Multi-scale Gate Mergence

Multi-modal magnetic resonance imaging (MRI) plays a critical role in clinical diagnosis and treatment nowadays. Each modality of MRI presents its own specific anatomical features which serve as complementary information to other modalities and can provide rich diagnostic information. However, due to the limitations of time consuming and expensive cost, some image sequences of patients may be lost or corrupted, posing an obstacle for accurate diagnosis. Although current multi-modal image synthesis approaches are able to alleviate the issues to some extent, they are still far short of fusing modalities effectively. In light of this, we propose a multi-scale gate mergence based generative adversarial network model, namely MGM-GAN, to synthesize one modality of MRI from others. Notably, we have multiple down-sampling branches corresponding to input modalities to specifically extract their unique features. In contrast to the generic multi-modal fusion approach of averaging or maximizing operations, we introduce a gate mergence (GM) mechanism to automatically learn the weights of different modalities across locations, enhancing the task-related information while suppressing the irrelative information. As such, the feature maps of all the input modalities at each down-sampling level, i.e., multi-scale levels, are integrated via GM module. In addition, both the adversarial loss and the pixel-wise loss, as well as gradient difference loss (GDL) are applied to train the network to produce the desired modality accurately. Extensive experiments demonstrate that the proposed method outperforms the state-of-the-art multi-modal image synthesis methods.

Applications of artificial intelligence in nuclear medicine image generation

Recently, the application of artificial intelligence (AI) in medical imaging (including nuclear medicine imaging) has rapidly developed. Most AI applications in nuclear medicine imaging have focused on the diagnosis, treatment monitoring, and correlation analyses with pathology or specific gene mutation. It can also be used for image generation to shorten the time of image acquisition, reduce the dose of injected tracer, and enhance image quality. This work provides an overview of the application of AI in image generation for single-photon emission computed tomography (SPECT) and positron emission tomography (PET) either without or with anatomical information [CT or magnetic resonance imaging (MRI)]. This review focused on four aspects, including imaging physics, image reconstruction, image postprocessing, and internal dosimetry. AI application in generating attenuation map, estimating scatter events, boosting image quality, and predicting internal dose map is summarized and discussed.

Monochromatic image reconstruction via machine learning

X-ray computed tomography (CT) is a nondestructive imaging technique to reconstruct cross-sectional images of an object using x-ray measurements taken from different view angles for medical diagnosis, therapeutic planning, security screening, and other applications. In clinical practice, the x-ray tube emits polychromatic x-rays, and the x-ray detector array operates in the energy-integrating mode to acquire energy intensity. This physical process of x-ray imaging is accurately described by an energy-dependent non-linear integral equation on the basis of the Beer–Lambert law. However, the non-linear model is not invertible using a computationally efficient solution and is often approximated as a linear integral model in the form of the Radon transform, which basically loses energy-dependent information. This approximate model produces an inaccurate quantification of attenuation images, suffering from beam-hardening effects. In this paper, a machine learning-based approach is proposed to correct the model mismatch to achieve quantitative CT imaging. Specifically, a one-dimensional network model is proposed to learn a non-linear transform from a training dataset to map a polychromatic CT image to its monochromatic sinogram at a pre-specified energy level, realizing virtual monochromatic (VM) imaging effectively and efficiently. Our results show that the proposed method recovers high-quality monochromatic projections with an average relative error of less than 2%. The resultant x-ray VM imaging can be applied for beam-hardening correction, material differentiation and tissue characterization, and proton therapy treatment planning.

Synthetic data in machine learning for medicine and healthcare

The proliferation of synthetic data in artificial intelligence for medicine and healthcare raises concerns about the vulnerabilities of the software and the challenges of current policy.

Adversarial Data Augmentation on Breast MRI Segmentation

The scarcity of balanced and annotated datasets has been a recurring problem in medical image analysis. Several researchers have tried to fill this gap employing dataset synthesis with adversarial networks (GANs). Breast magnetic resonance imaging (MRI) provides complex, texture-rich medical images, with the same annotation shortage issues, for which, to the best of our knowledge, no previous work tried synthesizing data. Within this context, our work addresses the problem of synthesizing breast MRI images from corresponding annotations and evaluate the impact of this data augmentation strategy on a semantic segmentation task. We explored variations of image-to-image translation using conditional GANs, namely fitting the generator’s architecture with residual blocks and experimenting with cycle consistency approaches. We studied the impact of these changes on visual verisimilarity and how an U-Net segmentation model is affected by the usage of synthetic data. We achieved sufficiently realistic-looking breast MRI images and maintained a stable segmentation score even when completely replacing the dataset with the synthetic set. Our results were promising, especially when concerning to Pix2PixHD and Residual CycleGAN architectures.

Three-dimensional self-attention conditional GAN with spectral normalization for multimodal neuroimaging synthesis

PURPOSE

To develop a new 3D generative adversarial network that is designed and optimized for the application of multimodal 3D neuroimaging synthesis.

METHODS

We present a 3D conditional generative adversarial network (GAN) that uses spectral normalization and feature matching to stabilize the training process and ensure optimization convergence (called SC-GAN). A self-attention module was also added to model the relationships between widely separated image voxels. The performance of the network was evaluated on the data set from ADNI-3, in which the proposed network was used to predict PET images, fractional anisotropy, and mean diffusivity maps from multimodal MRI. Then, SC-GAN was applied on a multidimensional diffusion MRI experiment for superresolution application. Experiment results were evaluated by normalized RMS error, peak SNR, and structural similarity.

RESULTS

In general, SC-GAN outperformed other state-of-the-art GAN networks including 3D conditional GAN in all three tasks across all evaluation metrics. Prediction error of the SC-GAN was 18%, 24% and 29% lower compared to 2D conditional GAN for fractional anisotropy, PET and mean diffusivity tasks, respectively. The ablation experiment showed that the major contributors to the improved performance of SC-GAN are the adversarial learning and the self-attention module, followed by the spectral normalization module. In the superresolution multidimensional diffusion experiment, SC-GAN provided superior predication in comparison to 3D Unet and 3D conditional GAN.

CONCLUSION

In this work, an efficient end-to-end framework for multimodal 3D medical image synthesis (SC-GAN) is presented. The source code is also made available at https://github.com/Haoyulance/SC-GAN.

Generation of annotated multimodal ground truth datasets for abdominal medical image registration

PURPOSE

Sparsity of annotated data is a major limitation in medical image processing tasks such as registration. Registered multimodal image data are essential for the diagnosis of medical conditions and the success of interventional medical procedures. To overcome the shortage of data, we present a method that allows the generation of annotated multimodal 4D datasets.

METHODS

We use a CycleGAN network architecture to generate multimodal synthetic data from the 4D extended cardiac-torso (XCAT) phantom and real patient data. Organ masks are provided by the XCAT phantom; therefore, the generated dataset can serve as ground truth for image segmentation and registration. Realistic simulation of respiration and heartbeat is possible within the XCAT framework. To underline the usability as a registration ground truth, a proof of principle registration is performed.

RESULTS

Compared to real patient data, the synthetic data showed good agreement regarding the image voxel intensity distribution and the noise characteristics. The generated T1-weighted magnetic resonance imaging, computed tomography (CT), and cone beam CT images are inherently co-registered. Thus, the synthetic dataset allowed us to optimize registration parameters of a multimodal non-rigid registration, utilizing liver organ masks for evaluation.

CONCLUSION

Our proposed framework provides not only annotated but also multimodal synthetic data which can serve as a ground truth for various tasks in medical imaging processing. We demonstrated the applicability of synthetic data for the development of multimodal medical image registration algorithms.

Research on obtaining pseudo CT images based on stacked generative adversarial network

Background

To investigate the feasibility of using a stacked generative adversarial network (sGAN) to synthesize pseudo computed tomography (CT) images based on ultrasound (US) images.

Methods

The pre-radiotherapy US and CT images of 75 patients with cervical cancer were selected for the training set of pseudo-image synthesis. In the first stage, labeled US images were used as the first conditional GAN input to obtain low-resolution pseudo CT images, and in the second stage, a super-resolution reconstruction GAN was used. The pseudo CT image obtained in the first stage was used as an input, following which a high-resolution pseudo CT image with clear texture and accurate grayscale information was obtained. Five cross validation tests were performed to verify our model. The mean absolute error (MAE) was used to compare each pseudo CT with the same patient's real CT image. Also, another 10 cases of patients with cervical cancer, before radiotherapy, were selected for testing, and the pseudo CT image obtained using the neural style transfer (NSF) and CycleGAN methods were compared with that obtained using the sGAN method proposed in this study. Finally, the dosimetric accuracy of pseudo CT images was verified by phantom experiments.

Results

The MAE metric values between the pseudo CT obtained based on sGAN, and the real CT in five-fold cross validation are 66.82±1.59 HU, 66.36±1.85 HU, 67.26±2.37 HU, 66.34±1.75 HU, and 67.22±1.30 HU, respectively. The results of the metrics, namely, normalized mutual information (NMI), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR), between the pseudo CT images obtained using the sGAN method and the ground truth CT (CTgt) images were compared with those of the other two methods via the paired t-test, and the differences were statistically significant. The dice similarity coefficient (DSC) measurement results showed that the pseudo CT images obtained using the sGAN method were more similar to the CTgt images of organs at risk. The dosimetric phantom experiments also showed that the dose distribution between the pseudo CT images synthesized by the new method was similar to that of the CTgt images.

Conclusions

Compared with NSF and CycleGAN methods, the sGAN method can obtain more accurate pseudo CT images, thereby providing a new method for image guidance in radiotherapy for cervical cancer.

Deep learning for in vivo near-infrared imaging

Detecting fluorescence in the second near-infrared window (NIR-II) up to ∼1,700 nm has emerged as a novel in vivo imaging modality with high spatial and temporal resolution through millimeter tissue depths. Imaging in the NIR-IIb window (1,500-1,700 nm) is the most effective one-photon approach to suppressing light scattering and maximizing imaging penetration depth, but relies on nanoparticle probes such as PbS/CdS containing toxic elements. On the other hand, imaging the NIR-I (700-1,000 nm) or NIR-IIa window (1,000-1,300 nm) can be done using biocompatible small-molecule fluorescent probes including US Food and Drug Administration-approved dyes such as indocyanine green (ICG), but has a caveat of suboptimal imaging quality due to light scattering. It is highly desired to achieve the performance of NIR-IIb imaging using molecular probes approved for human use. Here, we trained artificial neural networks to transform a fluorescence image in the shorter-wavelength NIR window of 900-1,300 nm (NIR-I/IIa) to an image resembling an NIR-IIb image. With deep-learning translation, in vivo lymph node imaging with ICG achieved an unprecedented signal-to-background ratio of >100. Using preclinical fluorophores such as IRDye-800, translation of ∼900-nm NIR molecular imaging of PD-L1 or EGFR greatly enhanced tumor-to-normal tissue ratio up to ∼20 from ∼5 and improved tumor margin localization. Further, deep learning greatly improved in vivo noninvasive NIR-II light-sheet microscopy (LSM) in resolution and signal/background. NIR imaging equipped with deep learning could facilitate basic biomedical research and empower clinical diagnostics and imaging-guided surgery in the clinic.

Franken-CT: Head and Neck MR-Based Pseudo-CT Synthesis Using Diverse Anatomical Overlapping MR-CT Scans

Typically, pseudo-Computerized Tomography (CT) synthesis schemes proposed in the literature rely on complete atlases acquired with the same field of view (FOV) as the input volume. However, clinical CTs are usually acquired in a reduced FOV to decrease patient ionization. In this work, we present the Franken-CT approach, showing how the use of a non-parametric atlas composed of diverse anatomical overlapping Magnetic Resonance (MR)-CT scans and deep learning methods based on the U-net architecture enable synthesizing extended head and neck pseudo-CTs. Visual inspection of the results shows the high quality of the pseudo-CT and the robustness of the method, which is able to capture the details of the bone contours despite synthesizing the resulting image from knowledge obtained from images acquired with a completely different FOV. The experimental Zero-Normalized Cross-Correlation (ZNCC) reports 0.9367 ± 0.0138 (mean ± SD) and 95% confidence interval (0.9221, 0.9512); the experimental Mean Absolute Error (MAE) reports 73.9149 ± 9.2101 HU and 95% confidence interval (66.3383, 81.4915); the Structural Similarity Index Measure (SSIM) reports 0.9943 ± 0.0009 and 95% confidence interval (0.9935, 0.9951); and the experimental Dice coefficient for bone tissue reports 0.7051 ± 0.1126 and 95% confidence interval (0.6125, 0.7977). The voxel-by-voxel correlation plot shows an excellent correlation between pseudo-CT and ground-truth CT Hounsfield Units (m = 0.87; adjusted R2 = 0.91; p < 0.001). The Bland–Altman plot shows that the average of the differences is low (−38.6471 ± 199.6100; 95% CI (−429.8827, 352.5884)). This work serves as a proof of concept to demonstrate the great potential of deep learning methods for pseudo-CT synthesis and their great potential using real clinical datasets.