Deep CT to MR Synthesis Using Paired and Unpaired Data

Synthetic CT generation based on multi-sequence MR using CycleGAN for head and neck MRI-only planning

The purpose of this study is to investigate the influence of different magnetic resonance (MR) sequences on the accuracy of generating computed tomography (sCT) images for nasopharyngeal carcinoma based on CycleGAN. In this study, 143 patients' head and neck MR sequence (T1, T2, T1C, and T1DIXONC) and CT imaging data were acquired. The generator and discriminator of CycleGAN are improved to achieve the purpose of balance confrontation, and a cyclic consistent structure control domain is proposed in terms of loss function. Four different single-sequence MR images and one multi-sequence MR image were used to evaluate the accuracy of sCT. During the model testing phase, five testing scenarios were employed to further assess the mean absolute error, peak signal-to-noise ratio, structural similarity index, and root mean square error between the actual CT images and the sCT images generated by different models. T1 sequence-based sCT achieved better results in single-sequence MR-based sCT. Multi-sequence MR-based sCT achieved better results with T1 sequence-based sCT in terms of evaluation metrics. For metrological evaluation, the global gamma passage rate of sCT based on sequence MR was greater than 95% at 3%/3 mm, except for sCT based on T2 sequence MR. We developed a CycleGAN method to synthesize CT using different MR sequences, this method shows encouraging potential for dosimetric evaluation.

MITS-GAN: Safeguarding medical imaging from tampering with generative adversarial networks

The progress in generative models, particularly Generative Adversarial Networks (GANs), opened new possibilities for image generation but raised concerns about potential malicious uses, especially in sensitive areas like medical imaging. This study introduces MITS-GAN, a novel approach to prevent tampering in medical images, with a specific focus on CT scans. The approach disrupts the output of the attacker's CT-GAN architecture by introducing finely tuned perturbations that are imperceptible to the human eye. Specifically, the proposed approach involves the introduction of appropriate Gaussian noise to the input as a protective measure against various attacks. Our method aims to enhance tamper resistance, comparing favorably to existing techniques. Experimental results on a CT scan demonstrate MITS-GAN's superior performance, emphasizing its ability to generate tamper-resistant images with negligible artifacts. As image tampering in medical domains poses life-threatening risks, our proactive approach contributes to the responsible and ethical use of generative models. This work provides a foundation for future research in countering cyber threats in medical imaging. Models and codes are publicly available.1.

An hetero-modal deep learning framework for medical image synthesis applied to contrast and non-contrast MRI

Some pathologies such as cancer and dementia require multiple imaging modalities to fully diagnose and assess the extent of the disease. Magnetic resonance imaging offers this kind of polyvalence, but examinations take time and can require contrast agent injection. The flexible synthesis of these imaging sequences based on the available ones for a given patient could help reduce scan times or circumvent the need for contrast agent injection. In this work, we propose a deep learning architecture that can perform the synthesis of all missing imaging sequences from any subset of available images. The network is trained adversarially, with the generator consisting of parallel 3D U-Net encoders and decoders that optimally combines their multi-resolution representations with a fusion operation learned by an attention network trained conjointly with the generator network. We compare our synthesis performance with 3D networks using other types of fusion and a comparable number of trainable parameters, such as the mean/variance fusion. In all synthesis scenarios except one, the synthesis performance of the network using attention-guided fusion was better than the other fusion schemes. We also inspect the encoded representations and the attention network outputs to gain insights into the synthesis process, and uncover desirable behaviors such as prioritization of specific modalities, flexible construction of the representation when important modalities are missing, and modalities being selected in regions where they carry sequence-specific information. This work suggests that a better construction of the latent representation space in hetero-modal networks can be achieved by using an attention network.

Feasibility study on the clinical application of CT-based synthetic brain T1-weighted MRI: comparison with conventional T1-weighted MRI

OBJECTIVES

This study aimed to examine the equivalence of computed tomography (CT)-based synthetic T1-weighted imaging (T1WI) to conventional T1WI for the quantitative assessment of brain morphology.

MATERIALS AND METHODS

This prospective study examined 35 adult patients undergoing brain magnetic resonance imaging (MRI) and CT scans. An image synthesis method based on a deep learning model was used to generate synthetic T1WI (sT1WI) from CT data. Two senior radiologists used sT1WI and conventional T1WI on separate occasions to independently measure clinically relevant brain morphological parameters. The reliability and consistency between conventional and synthetic T1WI were assessed using statistical consistency checks, comprising intra-reader, inter-reader, and inter-method agreement.

RESULTS

The intra-reader, inter-reader, and inter-method reliability and variability mostly exhibited the desired performance, except for several poor agreements due to measurement differences between the radiologists. All the measurements of sT1WI were equivalent to that of T1WI at 5% equivalent intervals.

CONCLUSION

This study demonstrated the equivalence of CT-based sT1WI to conventional T1WI for quantitatively assessing brain morphology, thereby providing more information on imaging diagnosis with a single CT scan.

CLINICAL RELEVANCE STATEMENT

Real-time synthesis of MR images from CT scans reduces the time required to acquire MR signals, improving the efficiency of the treatment planning system and providing benefits in the clinical diagnosis of patients with contraindications such as presence of metal implants or claustrophobia.

KEY POINTS

• Deep learning-based image synthesis methods generate synthetic T1-weighted imaging from CT scans. • The equivalence of synthetic T1-weighted imaging and conventional MRI for quantitative brain assessment was investigated. • Synthetic T1-weighted imaging can provide more information per scan and be used in preoperative diagnosis and radiotherapy.

Deep learning for the harmonization of structural MRI scans: a survey

Medical imaging datasets for research are frequently collected from multiple imaging centers using different scanners, protocols, and settings. These variations affect data consistency and compatibility across different sources. Image harmonization is a critical step to mitigate the effects of factors like inherent differences between various vendors, hardware upgrades, protocol changes, and scanner calibration drift, as well as to ensure consistent data for medical image processing techniques. Given the critical importance and widespread relevance of this issue, a vast array of image harmonization methodologies have emerged, with deep learning-based approaches driving substantial advancements in recent times. The goal of this review paper is to examine the latest deep learning techniques employed for image harmonization by analyzing cutting-edge architectural approaches in the field of medical image harmonization, evaluating both their strengths and limitations. This paper begins by providing a comprehensive fundamental overview of image harmonization strategies, covering three critical aspects: established imaging datasets, commonly used evaluation metrics, and characteristics of different scanners. Subsequently, this paper analyzes recent structural MRI (Magnetic Resonance Imaging) harmonization techniques based on network architecture, network learning algorithm, network supervision strategy, and network output. The underlying architectures include U-Net, Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), flow-based generative models, transformer-based approaches, as well as custom-designed network architectures. This paper investigates the effectiveness of Disentangled Representation Learning (DRL) as a pivotal learning algorithm in harmonization. Lastly, the review highlights the primary limitations in harmonization techniques, specifically the lack of comprehensive quantitative comparisons across different methods. The overall aim of this review is to serve as a guide for researchers and practitioners to select appropriate architectures based on their specific conditions and requirements. It also aims to foster discussions around ongoing challenges in the field and shed light on promising future research directions with the potential for significant advancements.

On-board synthetic 4D MRI generation from 4D CBCT for radiotherapy of abdominal tumors: A feasibility study

BACKGROUND

Magnetic resonance-guided radiotherapy with an MR-guided LINAC represents potential clinical benefits in abdominal treatments due to the superior soft-tissue contrast compared to kV-based images in conventional treatment units. However, due to the high cost associated with this technology, only a few centers have access to it. As an alternative, synthetic 4D MRI generation based on artificial intelligence methods could be implemented. Nevertheless, appropriate MRI texture generation from CT images might be challenging and prone to hallucinations, compromising motion accuracy.

PURPOSE

To evaluate the feasibility of on-board synthetic motion-resolved 4D MRI generation from prior 4D MRI, on-board 4D cone beam CT (CBCT) images, motion modeling information, and deep learning models using the digital anthropomorphic phantom XCAT.

METHODS

The synthetic 4D MRI corresponds to phases from on-board 4D CBCT. Each synthetic MRI volume in the 4D MRI was generated by warping a reference 3D MRI (MRIref, end of expiration phase from a prior 4D MRI) with a deformation field map (DFM) determined by (I) the eigenvectors from the principal component analysis (PCA) motion-modeling of the prior 4D MRI, and (II) the corresponding eigenvalues predicted by a convolutional neural network (CNN) model using the on-board 4D CBCT images as input. The CNN was trained with 1000 deformations of one reference CT (CTref, same conditions as MRIref) generated by applying 1000 DFMs computed by randomly sampling the original eigenvalues from the prior 4D MRI PCA model. The evaluation metrics for the CNN model were root-mean-square error (RMSE) and mean absolute error (MAE). Finally, different on-board 4D-MRI generation scenarios were assessed by changing the respiratory period, the amplitude of the diaphragm, and the chest wall motion of the 4D CBCT using normalized root-mean-square error (nRMSE) and structural similarity index measure (SSIM) for image-based evaluation, and volume dice coefficient (VDC), volume percent difference (VPD), and center-of-mass shift (COMS) for contour-based evaluation of liver and target volumes.

RESULTS

The RMSE and MAE values of the CNN model reported 0.012 ± 0.001 and 0.010 ± 0.001, respectively for the first eigenvalue predictions. SSIM and nRMSE were 0.96 ± 0.06 and 0.22 ± 0.08, respectively. VDC, VPD, and COMS were 0.92 ± 0.06, 3.08 ± 3.73 %, and 2.3 ± 2.1 mm, respectively, for the target volume. The more challenging synthetic 4D-MRI generation scenario was for one 4D-CBCT with increased chest wall motion amplitude, reporting SSIM and nRMSE of 0.82 and 0.51, respectively.

CONCLUSIONS

On-board synthetic 4D-MRI generation based on predicting actual treatment deformation from on-board 4D-CBCT represents a method that can potentially improve the treatment-setup localization in abdominal radiotherapy treatments with a conventional kV-based LINAC.

Deep Learning and Multimodal Artificial Intelligence in Orthopaedic Surgery

This review article focuses on the applications of deep learning with neural networks and multimodal neural networks in the orthopaedic domain. By providing practical examples of how artificial intelligence (AI) is being applied successfully in orthopaedic surgery, particularly in the realm of imaging data sets and the integration of clinical data, this study aims to provide orthopaedic surgeons with the necessary tools to not only evaluate existing literature but also to consider AI's potential in their own clinical or research pursuits. We first review standard deep neural networks which can analyze numerical clinical variables, then describe convolutional neural networks which can analyze image data, and then introduce multimodal AI models which analyze various types of different data. Then, we contrast these deep learning techniques with related but more limited techniques such as radiomics, describe how to interpret deep learning studies, and how to initiate such studies at your institution. Ultimately, by empowering orthopaedic surgeons with the knowledge and know-how of deep learning, this review aspires to facilitate the translation of research into clinical practice, thereby enhancing the efficacy and precision of real-world orthopaedic care for patients.

Multi-sequence generative adversarial network: better generation for enhanced magnetic resonance imaging images

Introduction

MRI is one of the commonly used diagnostic methods in clinical practice, especially in brain diseases. There are many sequences in MRI, but T1CE images can only be obtained by using contrast agents. Many patients (such as cancer patients) must undergo alignment of multiple MRI sequences for diagnosis, especially the contrast-enhanced magnetic resonance sequence. However, some patients such as pregnant women, children, etc. find it difficult to use contrast agents to obtain enhanced sequences, and contrast agents have many adverse reactions, which can pose a significant risk. With the continuous development of deep learning, the emergence of generative adversarial networks makes it possible to extract features from one type of image to generate another type of image.

Methods

We propose a generative adversarial network model with multimodal inputs and end-to-end decoding based on the pix2pix model. For the pix2pix model, we used four evaluation metrics: NMSE, RMSE, SSIM, and PNSR to assess the effectiveness of our generated model.

Results

Through statistical analysis, we compared our proposed new model with pix2pix and found significant differences between the two. Our model outperformed pix2pix, with higher SSIM and PNSR, lower NMSE and RMSE. We also found that the input of T1W images and T2W images had better effects than other combinations, providing new ideas for subsequent work on generating magnetic resonance enhancement sequence images. By using our model, it is possible to generate magnetic resonance enhanced sequence images based on magnetic resonance non-enhanced sequence images.

Discussion

This has significant implications as it can greatly reduce the use of contrast agents to protect populations such as pregnant women and children who are contraindicated for contrast agents. Additionally, contrast agents are relatively expensive, and this generation method may bring about substantial economic benefits.

Generative artificial intelligence in ophthalmology

Generative artificial intelligence (AI) has revolutionized medicine over the past several years. A generative adversarial network (GAN) is a deep learning framework that has become a powerful technique in medicine, particularly in ophthalmology for image analysis. In this paper we review the current ophthalmic literature involving GANs, and highlight key contributions in the field. We briefly touch on ChatGPT, another application of generative AI, and its potential in ophthalmology. We also explore the potential uses for GANs in ocular imaging, with a specific emphasis on 3 primary domains: image enhancement, disease identification, and generating of synthetic data. PubMed, Ovid MEDLINE, Google Scholar were searched from inception to October 30, 2022, to identify applications of GAN in ophthalmology. A total of 40 papers were included in this review. We cover various applications of GANs in ophthalmic-related imaging including optical coherence tomography, orbital magnetic resonance imaging, fundus photography, and ultrasound; however, we also highlight several challenges that resulted in the generation of inaccurate and atypical results during certain iterations. Finally, we examine future directions and considerations for generative AI in ophthalmology.

Image factory: A method for synthesizing novel CT images with anatomical guidance

BACKGROUND

Deep learning in medical applications is limited due to the low availability of large labeled, annotated, or segmented training datasets. With the insufficient data available for model training comes the inability of these networks to learn the fine nuances of the space of possible images in a given medical domain, leading to the possible suppression of important diagnostic features hence making these deep learning systems suboptimal in their performance and vulnerable to adversarial attacks.

PURPOSE

We formulate a framework to address this lack of labeled data problem. We test this formulation in computed tomographic images domain and present an approach that can synthesize large sets of novel CT images at high resolution across the full Hounsfield (HU) range.

METHODS

Our method only requires a small annotated dataset of lung CT from 30 patients (available online at the TCIA) and a large nonannotated dataset with high resolution CT images from 14k patients (received from NIH, not publicly available). It then converts the small annotated dataset into a large annotated dataset, using a sequence of steps including texture learning via StyleGAN, label learning via U-Net and semi-supervised learning via CycleGAN/Pixel-to-Pixel (P2P) architectures. The large annotated dataset so generated can then be used for the training of deep learning networks for medical applications. It can also be put to use for the synthesis of CT images with varied anatomies that were nonexistent within either of the input datasets, enriching the dataset even further.

RESULTS

We demonstrate our framework via lung CT-Scan synthesis along with their novel generated annotations and compared it with other state of the art generative models that only produce images without annotations. We evaluate our framework effectiveness via a visual turing test with help of a few doctors and radiologists.

CONCLUSIONS

We gain the capability of generating an unlimited amount of annotated CT images. Our approach works for all HU windows with minimal depreciation in anatomical plausibility and hence could be used as a general purpose framework for annotated data augmentation for deep learning applications in medical imaging.

One model to unite them all: Personalized federated learning of multi-contrast MRI synthesis

Curation of large, diverse MRI datasets via multi-institutional collaborations can help improve learning of generalizable synthesis models that reliably translate source- onto target-contrast images. To facilitate collaborations, federated learning (FL) adopts decentralized model training while mitigating privacy concerns by avoiding sharing of imaging data. However, conventional FL methods can be impaired by the inherent heterogeneity in the data distribution, with domain shifts evident within and across imaging sites. Here we introduce the first personalized FL method for MRI Synthesis (pFLSynth) that improves reliability against data heterogeneity via model specialization to individual sites and synthesis tasks (i.e., source-target contrasts). To do this, pFLSynth leverages an adversarial model equipped with novel personalization blocks that control the statistics of generated feature maps across the spatial/channel dimensions, given latent variables specific to sites and tasks. To further promote communication efficiency and site specialization, partial network aggregation is employed over later generator stages while earlier generator stages and the discriminator are trained locally. As such, pFLSynth enables multi-task training of multi-site synthesis models with high generalization performance across sites and tasks. Comprehensive experiments demonstrate the superior performance and reliability of pFLSynth in MRI synthesis against prior federated methods.

Turning brain MRI into diagnostic PET: 15O-water PET CBF synthesis from multi-contrast MRI via attention-based encoder-decoder networks

Accurate quantification of cerebral blood flow (CBF) is essential for the diagnosis and assessment of a wide range of neurological diseases. Positron emission tomography (PET) with radiolabeled water (15O-water) is the gold-standard for the measurement of CBF in humans, however, it is not widely available due to its prohibitive costs and the use of short-lived radiopharmaceutical tracers that require onsite cyclotron production. Magnetic resonance imaging (MRI), in contrast, is more accessible and does not involve ionizing radiation. This study presents a convolutional encoder-decoder network with attention mechanisms to predict the gold-standard 15O-water PET CBF from multi-contrast MRI scans, thus eliminating the need for radioactive tracers. The model was trained and validated using 5-fold cross-validation in a group of 126 subjects consisting of healthy controls and cerebrovascular disease patients, all of whom underwent simultaneous 15O-water PET/MRI. The results demonstrate that the model can successfully synthesize high-quality PET CBF measurements (with an average SSIM of 0.924 and PSNR of 38.8 dB) and is more accurate compared to concurrent and previous PET synthesis methods. We also demonstrate the clinical significance of the proposed algorithm by evaluating the agreement for identifying the vascular territories with impaired CBF. Such methods may enable more widespread and accurate CBF evaluation in larger cohorts who cannot undergo PET imaging due to radiation concerns, lack of access, or logistic challenges.

A systematic literature review: deep learning techniques for synthetic medical image generation and their applications in radiotherapy

The aim of this systematic review is to determine whether Deep Learning (DL) algorithms can provide a clinically feasible alternative to classic algorithms for synthetic Computer Tomography (sCT). The following categories are presented in this study: ∙ MR-based treatment planning and synthetic CT generation techniques. ∙ Generation of synthetic CT images based on Cone Beam CT images. ∙ Low-dose CT to High-dose CT generation. ∙ Attenuation correction for PET images. To perform appropriate database searches, we reviewed journal articles published between January 2018 and June 2023. Current methodology, study strategies, and results with relevant clinical applications were analyzed as we outlined the state-of-the-art of deep learning based approaches to inter-modality and intra-modality image synthesis. This was accomplished by contrasting the provided methodologies with traditional research approaches. The key contributions of each category were highlighted, specific challenges were identified, and accomplishments were summarized. As a final step, the statistics of all the cited works from various aspects were analyzed, which revealed that DL-based sCTs have achieved considerable popularity, while also showing the potential of this technology. In order to assess the clinical readiness of the presented methods, we examined the current status of DL-based sCT generation.

Synthesizing images of tau pathology from cross-modal neuroimaging using deep learning

Given the prevalence of dementia and the development of pathology-specific disease-modifying therapies, high-value biomarker strategies to inform medical decision-making are critical. In vivo tau-PET is an ideal target as a biomarker for Alzheimer's disease diagnosis and treatment outcome measure. However, tau-PET is not currently widely accessible to patients compared to other neuroimaging methods. In this study, we present a convolutional neural network (CNN) model that imputes tau-PET images from more widely available cross-modality imaging inputs. Participants (n = 1192) with brain T1-weighted MRI (T1w), fluorodeoxyglucose (FDG)-PET, amyloid-PET and tau-PET were included. We found that a CNN model can impute tau-PET images with high accuracy, the highest being for the FDG-based model followed by amyloid-PET and T1w. In testing implications of artificial intelligence-imputed tau-PET, only the FDG-based model showed a significant improvement of performance in classifying tau positivity and diagnostic groups compared to the original input data, suggesting that application of the model could enhance the utility of the metabolic images. The interpretability experiment revealed that the FDG- and T1w-based models utilized the non-local input from physically remote regions of interest to estimate the tau-PET, but this was not the case for the Pittsburgh compound B-based model. This implies that the model can learn the distinct biological relationship between FDG-PET, T1w and tau-PET from the relationship between amyloid-PET and tau-PET. Our study suggests that extending neuroimaging's use with artificial intelligence to predict protein specific pathologies has great potential to inform emerging care models.

Uncertainty-aware MR-based CT synthesis for robust proton therapy planning of brain tumour

BACKGROUND AND PURPOSE

Deep learning techniques excel in MR-based CT synthesis, but missing uncertainty prediction limits its clinical use in proton therapy. We developed an uncertainty-aware framework and evaluated its efficiency in robust proton planning.

MATERIALS AND METHODS

A conditional generative-adversarial network was trained on 64 brain tumour patients with paired MR-CT images to generate synthetic CTs (sCT) from combined T1-T2 MRs of three orthogonal planes. A Bayesian neural network predicts Laplacian distributions for all voxels with parameters (μ, b). A robust proton plan was optimized using three sCTs of μ and μ±b. The dosimetric differences between the plan from sCT (sPlan) and the recalculated plan (rPlan) on planning CT (pCT) were quantified for each patient. The uncertainty-aware robust plan was compared to conventional robust (global ± 3 %) and non-robust plans.

RESULTS

In 8-fold cross-validation, sCT-pCT image differences (Mean-Absolute-Error) were 80.84 ± 9.84HU (body), 35.78 ± 6.07HU (soft tissues) and 221.88 ± 31.69HU (bones), with Dice scores of 90.33 ± 2.43 %, 95.13 ± 0.80 %, and 85.53 ± 4.16 %, respectively. The uncertainty distribution positively correlated with absolute prediction error (Correlation Coefficient: 0.62 ± 0.01). The uncertainty-conditioned robust optimisation improved the rPlan-sPlan agreement, e.g., D95 absolute difference (CTV) was 1.10 ± 1.24 % compared to conventional (1.64 ± 2.71 %) and non-robust (2.08 ± 2.96 %) optimisation. This trend was consistent across all target and organs-at-risk indexes.

CONCLUSION

The enhanced framework incorporates 3D uncertainty prediction and generates high-quality sCTs from MR images. The framework also facilitates conditioned robust optimisation, bolstering proton plan robustness against network prediction errors. The innovative feature of uncertainty visualisation and robust analyses contribute to evaluating sCT clinical utility for individual patients.

CT-based generation of synthetic-pseudo MR images with different weightings for human knee

Synthetic MR images are generated for their high soft-tissue contrast avoiding the discomfort by the long acquisition time and placing claustrophobic patients in the MR scanner's confined space. The aim of this study is to generate synthetic pseudo-MR images from a real CT image for the knee region in vivo. 19 healthy subjects were scanned for model training, while 13 other healthy subjects were imaged for testing. The approach used in this work is novel such that the registration was performed between the MR and CT images, and the femur bone, patella, and the surrounding soft tissue were segmented on the CT image. The tissue type was mapped to its corresponding mean and standard deviation values of the CT# of a window moving on each pixel in the reconstructed CT images, which enabled the remapping of the tissue to its MRI intrinsic parameters: T1, T2, and proton density (ρ). To generate the synthetic MR image of a knee slice, a classic spin-echo sequence was simulated using proper intrinsic and contrast parameters. Results showed that the synthetic MR images were comparable to the real images acquired with the same TE and TR values, and the average slope between them (for all knee segments) was 0.98, while the average percentage root mean square difference (PRD) was 25.7%. In conclusion, this study has shown the feasibility and validity of accurately generating synthetic MR images of the knee region in vivo with different weightings from a single real CT image.

A deep learning approach to remove contrast from contrast-enhanced CT for proton dose calculation

PURPOSE

Non-Contrast Enhanced CT (NCECT) is normally required for proton dose calculation while Contrast Enhanced CT (CECT) is often scanned for tumor and organ delineation. Possible tissue motion between these two CTs raises dosimetry uncertainties, especially for moving tumors in the thorax and abdomen. Here we report a deep-learning approach to generate NCECT directly from CECT. This method could be useful to avoid the NCECT scan, reduce CT simulation time and imaging dose, and decrease the uncertainties caused by tissue motion between otherwise two different CT scans.

METHODS

A deep network was developed to convert CECT to NCECT. The network receives a 3D image from CECT images as input and generates a corresponding contrast-removed NCECT image patch. Abdominal CECT and NCECT image pairs of 20 patients were deformably registered and 8000 image patch pairs extracted from the registered image pairs were utilized to train and test the model. CTs of clinical proton patients and their treatment plans were employed to evaluate the dosimetric impact of using the generated NCECT for proton dose calculation.

RESULTS

Our approach achieved a Cosine Similarity score of 0.988 and an MSE value of 0.002. A quantitative comparison of clinical proton dose plans computed on the CECT and the generated NCECT for five proton patients revealed significant dose differences at the distal of beam paths. V100% of PTV and GTV changed by 3.5% and 5.5%, respectively. The mean HU difference for all five patients between the generated and the scanned NCECTs was ∼4.72, whereas the difference between CECT and the scanned NCECT was ∼64.52, indicating a ∼93% reduction in mean HU difference.

CONCLUSIONS

A deep learning approach was developed to generate NCECTs from CECTs. This approach could be useful for the proton dose calculation to reduce uncertainties caused by tissue motion between CECT and NCECT.

A qualitative study of improving megavoltage computed tomography image quality and maintaining dose accuracy using cycleGAN-based image synthesis

BACKGROUND

Due to inconsistent positioning, tumor shrinking, and weight loss during fractionated treatment, the initial plan was no longer appropriate after a few fractional treatments, and the patient will require adaptive helical tomotherapy (HT) to overcome the issue. Patients are scanned with megavoltage computed tomography (MVCT) before each fractional treatment, which is utilized for patient setup and provides information for dose reconstruction. However, the low contrast and high noise of MVCT make it challenging to delineate treatment targets and organs at risk (OAR).

PURPOSE

This study developed a deep-learning-based approach to generate high-quality synthetic kilovoltage computed tomography (skVCT) from MVCT and meet clinical dose requirements.

METHODS

Data from 41 head and neck cancer patients were collected; 25 (2995 slices) were used for training, and 16 (1898 slices) for testing. A cycle generative adversarial network (cycleGAN) based on attention gate and residual blocks was used to generate MVCT-based skVCT. For the 16 patients, kVCT-based plans were transferred to skVCT images and electron density profile-corrected MVCT images to recalculate the dose. The quantitative indices and clinically relevant dosimetric metrics, including the mean absolute error (MAE), structural similarity index measure (SSIM), peak signal-to-noise ratio (PSNR), gamma passing rates, and dose-volume-histogram (DVH) parameters (Dmax , Dmean , Dmin ), were used to assess the skVCT images.

RESULTS

The MAE, PSNR, and SSIM of MVCT were 109.6 ± 12.3 HU, 27.5 ± 1.1 dB, and 91.9% ± 1.7%, respectively, while those of skVCT were 60.6 ± 9.0 HU, 34.0 ± 1.9 dB, and 96.5% ± 1.1%. The image quality and contrast were enhanced, and the noise was reduced. The gamma passing rates improved from 98.31% ± 1.11% to 99.71% ± 0.20% (2 mm/2%) and 99.77% ± 0.18% to 99.98% ± 0.02% (3 mm/3%). No significant differences (p > 0.05) were observed in DVH parameters between kVCT and skVCT.

CONCLUSION

With training on a small data set (2995 slices), the model successfully generated skVCT with improved image quality, and the dose calculation accuracy was similar to that of MVCT. MVCT-based skVCT can increase treatment accuracy and offer the possibility of implementing adaptive radiotherapy.

Unsupervised synthesis of realistic coronary artery X-ray angiogram

PURPOSE

Medical image analysis suffers from a sparsity of annotated data necessary in learning-based models. Cardiorespiratory simulators have been developed to counter the lack of data. However, the resulting data often lack realism. Hence, the proposed method aims to synthesize realistic and fully customizable angiograms of coronary arteries for the training of learning-based biomedical tasks, for cardiologists performing interventions, and for cardiologist trainees.

METHODS

3D models of coronary arteries are generated with a fully customizable realistic cardiorespiratory simulator. The transfer of X-ray angiography style to simulator-generated images is performed using a new vessel-specific adaptation of the CycleGAN model. The CycleGAN model is paired with a vesselness-based loss function that is designed as a vessel-specific structural integrity constraint.

RESULTS

Validation is performed both on the style and on the preservation of the shape of the arteries of the images. The results show a PSNR of 14.125, an SSIM of 0.898, and an overlapping of 89.5% using the Dice coefficient.

CONCLUSION

We proposed a novel fluoroscopy-based style transfer method for the enhancement of the realism of simulated coronary artery angiograms. The results show that the proposed model is capable of accurately transferring the style of X-ray angiograms to the simulations while keeping the integrity of the structures of interest (i.e., the topology of the coronary arteries).

Unsupervised Medical Image Translation With Adversarial Diffusion Models

Imputation of missing images via source-to-target modality translation can improve diversity in medical imaging protocols. A pervasive approach for synthesizing target images involves one-shot mapping through generative adversarial networks (GAN). Yet, GAN models that implicitly characterize the image distribution can suffer from limited sample fidelity. Here, we propose a novel method based on adversarial diffusion modeling, SynDiff, for improved performance in medical image translation. To capture a direct correlate of the image distribution, SynDiff leverages a conditional diffusion process that progressively maps noise and source images onto the target image. For fast and accurate image sampling during inference, large diffusion steps are taken with adversarial projections in the reverse diffusion direction. To enable training on unpaired datasets, a cycle-consistent architecture is devised with coupled diffusive and non-diffusive modules that bilaterally translate between two modalities. Extensive assessments are reported on the utility of SynDiff against competing GAN and diffusion models in multi-contrast MRI and MRI-CT translation. Our demonstrations indicate that SynDiff offers quantitatively and qualitatively superior performance against competing baselines.

Artificial CT images can enhance variation of case images in diagnostic radiology skills training

OBJECTIVES

We sought to investigate if artificial medical images can blend with original ones and whether they adhere to the variable anatomical constraints provided.

METHODS

Artificial images were generated with a generative model trained on publicly available standard and low-dose chest CT images (805 scans; 39,803 2D images), of which 17% contained evidence of pathological formations (lung nodules). The test set (90 scans; 5121 2D images) was used to assess if artificial images (512 × 512 primary and control image sets) blended in with original images, using both quantitative metrics and expert opinion. We further assessed if pathology characteristics in the artificial images can be manipulated.

RESULTS

Primary and control artificial images attained an average objective similarity of 0.78 ± 0.04 (ranging from 0 [entirely dissimilar] to 1[identical]) and 0.76 ± 0.06, respectively. Five radiologists with experience in chest and thoracic imaging provided a subjective measure of image quality; they rated artificial images as 3.13 ± 0.46 (range of 1 [unrealistic] to 4 [almost indistinguishable to the original image]), close to their rating of the original images (3.73 ± 0.31). Radiologists clearly distinguished images in the control sets (2.32 ± 0.48 and 1.07 ± 0.19). In almost a quarter of the scenarios, they were not able to distinguish primary artificial images from the original ones.

CONCLUSION

Artificial images can be generated in a way such that they blend in with original images and adhere to anatomical constraints, which can be manipulated to augment the variability of cases.

CRITICAL RELEVANCE STATEMENT

Artificial medical images can be used to enhance the availability and variety of medical training images by creating new but comparable images that can blend in with original images.

KEY POINTS

• Artificial images, similar to original ones, can be created using generative networks. • Pathological features of artificial images can be adjusted through guiding the network. • Artificial images proved viable to augment the depth and broadening of diagnostic training.

A Survey on Artificial Intelligence in Pulmonary Imaging

Over the last decade, deep learning (DL) has contributed a paradigm shift in computer vision and image recognition creating widespread opportunities of using artificial intelligence in research as well as industrial applications. DL has been extensively studied in medical imaging applications, including those related to pulmonary diseases. Chronic obstructive pulmonary disease, asthma, lung cancer, pneumonia, and, more recently, COVID-19 are common lung diseases affecting nearly 7.4% of world population. Pulmonary imaging has been widely investigated toward improving our understanding of disease etiologies and early diagnosis and assessment of disease progression and clinical outcomes. DL has been broadly applied to solve various pulmonary image processing challenges including classification, recognition, registration, and segmentation. This paper presents a survey of pulmonary diseases, roles of imaging in translational and clinical pulmonary research, and applications of different DL architectures and methods in pulmonary imaging with emphasis on DL-based segmentation of major pulmonary anatomies such as lung volumes, lung lobes, pulmonary vessels, and airways as well as thoracic musculoskeletal anatomies related to pulmonary diseases.

Deep-Learning-Aided Evaluation of Spondylolysis Imaged with Ultrashort Echo Time Magnetic Resonance Imaging

Isthmic spondylolysis results in fracture of pars interarticularis of the lumbar spine, found in as many as half of adolescent athletes with persistent low back pain. While computed tomography (CT) is the gold standard for the diagnosis of spondylolysis, the use of ionizing radiation near reproductive organs in young subjects is undesirable. While magnetic resonance imaging (MRI) is preferable, it has lowered sensitivity for detecting the condition. Recently, it has been shown that ultrashort echo time (UTE) MRI can provide markedly improved bone contrast compared to conventional MRI. To take UTE MRI further, we developed supervised deep learning tools to generate (1) CT-like images and (2) saliency maps of fracture probability from UTE MRI, using ex vivo preparation of cadaveric spines. We further compared quantitative metrics of the contrast-to-noise ratio (CNR), mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) between UTE MRI (inverted to make the appearance similar to CT) and CT and between CT-like images and CT. Qualitative results demonstrated the feasibility of successfully generating CT-like images from UTE MRI to provide easier interpretability for bone fractures thanks to improved image contrast and CNR. Quantitatively, the mean CNR of bone against defect-filled tissue was 35, 97, and 146 for UTE MRI, CT-like, and CT images, respectively, being significantly higher for CT-like than UTE MRI images. For the image similarity metrics using the CT image as the reference, CT-like images provided a significantly lower mean MSE (0.038 vs. 0.0528), higher mean PSNR (28.6 vs. 16.5), and higher SSIM (0.73 vs. 0.68) compared to UTE MRI images. Additionally, the saliency maps enabled quick detection of the location with probable pars fracture by providing visual cues to the reader. This proof-of-concept study is limited to the data from ex vivo samples, and additional work in human subjects with spondylolysis would be necessary to refine the models for clinical use. Nonetheless, this study shows that the utilization of UTE MRI and deep learning tools could be highly useful for the evaluation of isthmic spondylolysis.

Dosimetric and workflow impact of synthetic-MRI use in prostate high-dose-rate brachytherapy

PURPOSE

Target and organ delineation during prostate high-dose-rate (HDR) brachytherapy treatment planning can be improved by acquiring both a postimplant CT and MRI. However, this leads to a longer treatment delivery workflow and may introduce uncertainties due to anatomical motion between scans. We investigated the dosimetric and workflow impact of MRI synthesized from CT for prostate HDR brachytherapy.

METHODS AND MATERIALS

Seventy-eight CT and T2-weighted MRI datasets from patients treated with prostate HDR brachytherapy at our institution were retrospectively collected to train and validate our deep-learning-based image-synthesis method. Synthetic MRI was assessed against real MRI using the dice similarity coefficient (DSC) between prostate contours drawn using both image sets. The DSC between the same observer's synthetic and real MRI prostate contours was compared with the DSC between two different observers' real MRI prostate contours. New treatment plans were generated targeting the synthetic MRI-defined prostate and compared with the clinically delivered plans using target coverage and dose to critical organs.

RESULTS

Variability between the same observer's prostate contours from synthetic and real MRI was not significantly different from the variability between different observer's prostate contours on real MRI. Synthetic MRI-planned target coverage was not significantly different from that of the clinically delivered plans. There were no increases above organ institutional dose constraints in the synthetic MRI plans.

CONCLUSIONS

We developed and validated a method for synthesizing MRI from CT for prostate HDR brachytherapy treatment planning. Synthetic MRI use may lead to a workflow advantage and removal of CT-to-MRI registration uncertainty without loss of information needed for target delineation and treatment planning.

Within-Modality Synthesis and Novel Radiomic Evaluation of Brain MRI Scans

One of the most common challenges in brain MRI scans is to perform different MRI sequences depending on the type and properties of tissues. In this paper, we propose a generative method to translate T2-Weighted (T2W) Magnetic Resonance Imaging (MRI) volume from T2-weight-Fluid-attenuated-Inversion-Recovery (FLAIR) and vice versa using Generative Adversarial Networks (GAN). To evaluate the proposed method, we propose a novel evaluation schema for generative and synthetic approaches based on radiomic features. For the evaluation purpose, we consider 510 pair-slices from 102 patients to train two different GAN-based architectures Cycle GAN and Dual Cycle-Consistent Adversarial network (DC2Anet). The results indicate that generative methods can produce similar results to the original sequence without significant change in the radiometric feature. Therefore, such a method can assist clinics to make decisions based on the generated image when different sequences are not available or there is not enough time to re-perform the MRI scans.

CycleGAN with mutual information loss constraint generates structurally aligned CT images from functional EIT images

Electrical impedance tomography (EIT) has been employed in the field of medical imaging due to its cost effectiveness, safety profile and portability, but the images generated are relatively low resolution. To address these limitations, we create a novel method using EIT images to generate high resolution structurally aligned images of lungs like those from CT scans. A way to achieve this transformation is via Cycle generative adversarial networks (CycleGAN), which have demonstrated image-to-image translation capabilities across different modalities. However, a generic implementation yields images which may not be aligned with their input image. To solve this issue, we construct and incorporate a Mutual Information (MI) constraint in CycleGAN to translate functional lung EIT images to structural high resolution CT images. The CycleGAN is first trained on unpaired EIT and CT lung images. Afterwards, we generate CT image pairs from EIT images via CycleGANs constrained with MI loss and without this loss. Finally, through generating these 1560 CT image pairs and then comparing the visual results and quantitative metrics, we show that MI constrained CycleGAN produces more structurally aligned CT images, where Normalised Mutual Information (NMI) is increased to 0.2621+/- 0.0052 versus 0.2600 +/- 0.0066, p<0.0001 for non-MI constrained images. By this process, we simultaneously provide functional and structural information, and potentially enable more detailed assessment of lungs.Clinical Relevance- By establishing a structurally aligning generative process via MI Loss in CycleGAN, this study enables EIT-CT conversion, thereby providing functional and structural images for enhanced lung assessment, from just EIT images.

Synthetic computed tomography for low-field magnetic resonance-only radiotherapy in head-and-neck cancer using residual vision transformers

Background and purpose

Synthetic computed tomography (sCT) scans are necessary for dose calculation in magnetic resonance (MR)-only radiotherapy. While deep learning (DL) has shown remarkable performance in generating sCT scans from MR images, research has predominantly focused on high-field MR images. This study presents the first implementation of a DL model for sCT generation in head-and-neck (HN) cancer using low-field MR images. Specifically, the use of vision transformers (ViTs) was explored.

Materials and methods

The dataset consisted of 31 patients, resulting in 196 pairs of deformably-registered computed tomography (dCT) and MR scans. The latter were obtained using a balanced steady-state precession sequence on a 0.35T scanner. Residual ViTs were trained on 2D axial, sagittal, and coronal slices, respectively, and the final sCTs were generated by averaging the models' outputs. Different image similarity metrics, dose volume histogram (DVH) deviations, and gamma analyses were computed on the test set (n = 6). The overlap between auto-contours on sCT scans and manual contours on MR images was evaluated for different organs-at-risk using the Dice score.

Results

The median [range] value of the test mean absolute error was 57 [37-74] HU. DVH deviations were below 1% for all structures. The median gamma passing rates exceeded 94% in the 2%/2mm analysis (threshold = 90%). The median Dice scores were above 0.7 for all organs-at-risk.

Conclusions

The clinical applicability of DL-based sCT generation from low-field MR images in HN cancer was proved. High sCT-dCT similarity and dose metric accuracy were achieved, and sCT suitability for organs-at-risk auto-delineation was shown.

The effect of neural network architecture on virtual H&E staining: Systematic assessment of histological feasibility

Conventional histopathology has relied on chemical staining for over a century. The staining process makes tissue sections visible to the human eye through a tedious and labor-intensive procedure that alters the tissue irreversibly, preventing repeated use of the sample. Deep learning-based virtual staining can potentially alleviate these shortcomings. Here, we used standard brightfield microscopy on unstained tissue sections and studied the impact of increased network capacity on the resulting virtually stained H&E images. Using the generative adversarial neural network model pix2pix as a baseline, we observed that replacing simple convolutions with dense convolution units increased the structural similarity score, peak signal-to-noise ratio, and nuclei reproduction accuracy. We also demonstrated highly accurate reproduction of histology, especially with increased network capacity, and demonstrated applicability to several tissues. We show that network architecture optimization can improve the image translation accuracy of virtual H&E staining, highlighting the potential of virtual staining in streamlining histopathological analysis.

Cross-modality image translation: CT image synthesis of MR brain images using multi generative network with perceptual supervision

BACKGROUND

Computed tomography (CT) and magnetic resonance imaging (MRI) are the mainstream imaging technologies for clinical practice. CT imaging can reveal high-quality anatomical and physiopathological structures, especially bone tissue, for clinical diagnosis. MRI provides high resolution in soft tissue and is sensitive to lesions. CT combined with MRI diagnosis has become a regular image-guided radiation treatment plan.

METHODS

In this paper, to reduce the dose of radiation exposure in CT examinations and ameliorate the limitations of traditional virtual imaging technologies, we propose a Generative MRI-to-CT transformation method with structural perceptual supervision. Even though structural reconstruction is structurally misaligned in the MRI-CT dataset registration, our proposed method can better align structural information of synthetic CT (sCT) images to input MRI images while simulating the modality of CT in the MRI-to-CT cross-modality transformation.

RESULTS

We retrieved a total of 3416 brain MRI-CT paired images as the train/test dataset, including 1366 train images of 10 patients and 2050 test images of 15 patients. Several methods (the baseline methods and the proposed method) were evaluated by the HU difference map, HU distribution, and various similarity metrics, including the mean absolute error (MAE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and normalized cross-correlation (NCC). In our quantitative experimental results, the proposed method achieves the lowest MAE mean of 0.147, highest PSNR mean of 19.27, and NCC mean of 0.431 in the overall CT test dataset.

CONCLUSIONS

In conclusion, both qualitative and quantitative results of synthetic CT validate that the proposed method can preserve higher similarity of structural information of the bone tissue of target CT than the baseline methods. Furthermore, the proposed method provides better HU intensity reconstruction for simulating the distribution of the CT modality. The experimental estimation indicates that the proposed method is worth further investigation.

QUANTITATIVE BIOMARKERS REPRODUCIBILITY USING GENERATIVE ADVERSARIAL APPROACHES IN REDUCED TO CONVENTIONAL DOSE CT

In recent years, several techniques for image-to-image translation by means of generative adversarial neural networks (GAN) have been proposed to learn mapping characteristics between a source and a target domain. In particular, in the medical imaging field conditional GAN frameworks with paired samples (cGAN) and unconditional cycle-consistent GANs with unpaired data (CycleGAN) have been demonstrated as a powerful scheme to model non-linear mappings that produce realistic target images from different modality sources. When proposing the usage and adaptation of these frameworks for medical image synthesis, quantitative and qualitative validation are usually performed by assessing the similarity between synthetic and target images in terms of metrics such as mean absolute error (MAE) or structural similarity (SSIM) index. However, an evaluation of clinically relevant markers showing that diagnostic information is not overlooked in the translation process is often missing. In this work, we aim at demonstrating the importance of validating medical image-to-image translation techniques by assessing their effect on the measurement of clinically relevant metrics and biomarkers. We implemented both a conditional and an unconditional approach to synthesize conventional dose chest CT scans from reduced dose CT and show that while both visually and in terms of traditional metrics the network appears to successfully minimize perceptual discrepancies, these methods are not reliable to systematically reproduce quantitative measurements of various chest biomarkers.

Deep Learning Subtraction Angiography: Improved Generalizability with Transfer Learning

PURPOSE

To investigate the utility and generalizability of deep learning subtraction angiography (DLSA) for generating synthetic digital subtraction angiography (DSA) images without misalignment artifacts.

MATERIALS AND METHODS

DSA images and native digital angiograms of the cerebral, hepatic, and splenic vasculature, both with and without motion artifacts, were retrospectively collected. Images were divided into a motion-free training set (n = 66 patients, 9,161 images) and a motion artifact-containing test set (n = 22 patients, 3,322 images). Using the motion-free set, the deep neural network pix2pix was trained to produce synthetic DSA images without misalignment artifacts directly from native digital angiograms. After training, the algorithm was tested on digital angiograms of hepatic and splenic vasculature with substantial motion. Four board-certified radiologists evaluated performance via visual assessment using a 5-grade Likert scale. Subgroup analyses were performed to analyze the impact of transfer learning and generalizability to novel vasculature.

RESULTS

Compared with the traditional DSA method, the proposed approach was found to generate synthetic DSA images with significantly fewer background artifacts (a mean rating of 1.9 [95% CI, 1.1-2.6] vs 3.5 [3.5-4.4]; P = .01) without a significant difference in foreground vascular detail (mean rating of 3.1 [2.6-3.5] vs 3.3 [2.8-3.8], P = .19) in both the hepatic and splenic vasculature. Transfer learning significantly improved the quality of generated images (P < .001).

CONCLUSIONS

DLSA successfully generates synthetic angiograms without misalignment artifacts, is improved through transfer learning, and generalizes reliably to novel vasculature that was not included in the training data.

Ensemble learning and personalized training for the improvement of unsupervised deep learning-based synthetic CT reconstruction

BACKGROUND

The growing adoption of magnetic resonance imaging (MRI)-guided radiation therapy (RT) platforms and a focus on MRI-only RT workflows have brought the technical challenge of synthetic computed tomography (sCT) reconstruction to the forefront. Unpaired-data deep learning-based approaches to the problem offer the attractive characteristic of not requiring paired training data, but the gap between paired- and unpaired-data results can be limiting.

PURPOSE

We present two distinct approaches aimed at improving unpaired-data sCT reconstruction results: a cascade ensemble that combines multiple models and a personalized training strategy originally designed for the paired-data setting.

METHODS

Comparisons are made between the following models: (1) the paired-data fully convolutional DenseNet (FCDN), (2) the FCDN with the Intentional Deep Overfit Learning (IDOL) personalized training strategy, (3) the unpaired-data CycleGAN, (4) the CycleGAN with the IDOL training strategy, and (5) the CycleGAN as an intermediate model in a cascade ensemble approach. Evaluation of the various models over 25 total patients is carried out using a five-fold cross-validation scheme, with the patient-specific IDOL models being trained for the five patients of fold 3, chosen at random.

RESULTS

In both the paired- and unpaired-data settings, adopting the IDOL training strategy led to improvements in the mean absolute error (MAE) between true CT images and sCT outputs within the body contour (mean improvement, paired- and unpaired-data approaches, respectively: 38%, 9%) and in regions of bone (52%, 5%), the peak signal-to-noise ratio (PSNR; 15%, 7%), and the structural similarity index (SSIM; 6%, <1%). The ensemble approach offered additional benefits over the IDOL approach in all three metrics (mean improvement over unpaired-data approach in fold 3; MAE: 20%; bone MAE: 16%; PSNR: 10%; SSIM: 2%), and differences in body MAE between the ensemble approach and the paired-data approach are statistically insignificant.

CONCLUSIONS

We have demonstrated that both a cascade ensemble approach and a personalized training strategy designed initially for the paired-data setting offer significant improvements in image quality metrics for the unpaired-data sCT reconstruction task. Closing the gap between paired- and unpaired-data approaches is a step toward fully enabling these powerful and attractive unpaired-data frameworks.

Deep learning based unpaired image-to-image translation applications for medical physics: a systematic review

Purpose. There is a growing number of publications on the application of unpaired image-to-image (I2I) translation in medical imaging. However, a systematic review covering the current state of this topic for medical physicists is lacking. The aim of this article is to provide a comprehensive review of current challenges and opportunities for medical physicists and engineers to apply I2I translation in practice.Methods and materials. The PubMed electronic database was searched using terms referring to unpaired (unsupervised), I2I translation, and medical imaging. This review has been reported in compliance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. From each full-text article, we extracted information extracted regarding technical and clinical applications of methods, Transparent Reporting for Individual Prognosis Or Diagnosis (TRIPOD) study type, performance of algorithm and accessibility of source code and pre-trained models.Results. Among 461 unique records, 55 full-text articles were included in the review. The major technical applications described in the selected literature are segmentation (26 studies), unpaired domain adaptation (18 studies), and denoising (8 studies). In terms of clinical applications, unpaired I2I translation has been used for automatic contouring of regions of interest in MRI, CT, x-ray and ultrasound images, fast MRI or low dose CT imaging, CT or MRI only based radiotherapy planning, etc Only 5 studies validated their models using an independent test set and none were externally validated by independent researchers. Finally, 12 articles published their source code and only one study published their pre-trained models.Conclusion. I2I translation of medical images offers a range of valuable applications for medical physicists. However, the scarcity of external validation studies of I2I models and the shortage of publicly available pre-trained models limits the immediate applicability of the proposed methods in practice.

VSGD-Net: Virtual Staining Guided Melanocyte Detection on Histopathological Images

Detection of melanocytes serves as a critical prerequisite in assessing melanocytic growth patterns when diagnosing melanoma and its precursor lesions on skin biopsy specimens. However, this detection is challenging due to the visual similarity of melanocytes to other cells in routine Hematoxylin and Eosin (H&E) stained images, leading to the failure of current nuclei detection methods. Stains such as Sox10 can mark melanocytes, but they require an additional step and expense and thus are not regularly used in clinical practice. To address these limitations, we introduce VSGD-Net, a novel detection network that learns melanocyte identification through virtual staining from H&E to Sox10. The method takes only routine H&E images during inference, resulting in a promising approach to support pathologists in the diagnosis of melanoma. To the best of our knowledge, this is the first study that investigates the detection problem using image synthesis features between two distinct pathology stainings. Extensive experimental results show that our proposed model outperforms state-of-the-art nuclei detection methods for melanocyte detection. The source code and pre-trained model are available at: https://github.com/kechunl/VSGD-Net.

Semi-Supervised Learning of MRI Synthesis Without Fully-Sampled Ground Truths

Learning-based translation between MRI contrasts involves supervised deep models trained using high-quality source- and target-contrast images derived from fully-sampled acquisitions, which might be difficult to collect under limitations on scan costs or time. To facilitate curation of training sets, here we introduce the first semi-supervised model for MRI contrast translation (ssGAN) that can be trained directly using undersampled k-space data. To enable semi-supervised learning on undersampled data, ssGAN introduces novel multi-coil losses in image, k-space, and adversarial domains. The multi-coil losses are selectively enforced on acquired k-space samples unlike traditional losses in single-coil synthesis models. Comprehensive experiments on retrospectively undersampled multi-contrast brain MRI datasets are provided. Our results demonstrate that ssGAN yields on par performance to a supervised model, while outperforming single-coil models trained on coil-combined magnitude images. It also outperforms cascaded reconstruction-synthesis models where a supervised synthesis model is trained following self-supervised reconstruction of undersampled data. Thus, ssGAN holds great promise to improve the feasibility of learning-based multi-contrast MRI synthesis.

Shifting machine learning for healthcare from development to deployment and from models to data

In the past decade, the application of machine learning (ML) to healthcare has helped drive the automation of physician tasks as well as enhancements in clinical capabilities and access to care. This progress has emphasized that, from model development to model deployment, data play central roles. In this Review, we provide a data-centric view of the innovations and challenges that are defining ML for healthcare. We discuss deep generative models and federated learning as strategies to augment datasets for improved model performance, as well as the use of the more recent transformer models for handling larger datasets and enhancing the modelling of clinical text. We also discuss data-focused problems in the deployment of ML, emphasizing the need to efficiently deliver data to ML models for timely clinical predictions and to account for natural data shifts that can deteriorate model performance.

Deep-learning-based generation of synthetic 6-minute MRI from 2-minute MRI for use in head and neck cancer radiotherapy

Background

Quick magnetic resonance imaging (MRI) scans with low contrast-to-noise ratio are typically acquired for daily MRI-guided radiotherapy setup. However, for patients with head and neck (HN) cancer, these images are often insufficient for discriminating target volumes and organs at risk (OARs). In this study, we investigated a deep learning (DL) approach to generate high-quality synthetic images from low-quality images.

Methods

We used 108 unique HN image sets of paired 2-minute T2-weighted scans (2mMRI) and 6-minute T2-weighted scans (6mMRI). 90 image sets (~20,000 slices) were used to train a 2-dimensional generative adversarial DL model that utilized 2mMRI as input and 6mMRI as output. Eighteen image sets were used to test model performance. Similarity metrics, including the mean squared error (MSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) were calculated between normalized synthetic 6mMRI and ground-truth 6mMRI for all test cases. In addition, a previously trained OAR DL auto-segmentation model was used to segment the right parotid gland, left parotid gland, and mandible on all test case images. Dice similarity coefficients (DSC) were calculated between 2mMRI and either ground-truth 6mMRI or synthetic 6mMRI for each OAR; two one-sided t-tests were applied between the ground-truth and synthetic 6mMRI to determine equivalence. Finally, a visual Turing test using paired ground-truth and synthetic 6mMRI was performed using three clinician observers; the percentage of images that were correctly identified was compared to random chance using proportion equivalence tests.

Results

The median similarity metrics across the whole images were 0.19, 0.93, and 33.14 for MSE, SSIM, and PSNR, respectively. The median of DSCs comparing ground-truth vs. synthetic 6mMRI auto-segmented OARs were 0.86 vs. 0.85, 0.84 vs. 0.84, and 0.82 vs. 0.85 for the right parotid gland, left parotid gland, and mandible, respectively (equivalence p<0.05 for all OARs). The percent of images correctly identified was equivalent to chance (p<0.05 for all observers).

Conclusions

Using 2mMRI inputs, we demonstrate that DL-generated synthetic 6mMRI outputs have high similarity to ground-truth 6mMRI, but further improvements can be made. Our study facilitates the clinical incorporation of synthetic MRI in MRI-guided radiotherapy.

Medical image translation using an edge-guided generative adversarial network with global-to-local feature fusion

In this paper, we propose a framework based deep learning for medical image translation using paired and unpaired training data. Initially, a deep neural network with an encoder-decoder structure is proposed for image-to-image translation using paired training data. A multi-scale context aggregation approach is then used to extract various features from different levels of encoding, which are used during the corresponding network decoding stage. At this point, we further propose an edge-guided generative adversarial network for image-to-image translation based on unpaired training data. An edge constraint loss function is used to improve network performance in tissue boundaries. To analyze framework performance, we conducted five different medical image translation tasks. The assessment demonstrates that the proposed deep learning framework brings significant improvement beyond state-of-the-arts.

ResViT: Residual Vision Transformers for Multimodal Medical Image Synthesis

Generative adversarial models with convolutional neural network (CNN) backbones have recently been established as state-of-the-art in numerous medical image synthesis tasks. However, CNNs are designed to perform local processing with compact filters, and this inductive bias compromises learning of contextual features. Here, we propose a novel generative adversarial approach for medical image synthesis, ResViT, that leverages the contextual sensitivity of vision transformers along with the precision of convolution operators and realism of adversarial learning. ResViT's generator employs a central bottleneck comprising novel aggregated residual transformer (ART) blocks that synergistically combine residual convolutional and transformer modules. Residual connections in ART blocks promote diversity in captured representations, while a channel compression module distills task-relevant information. A weight sharing strategy is introduced among ART blocks to mitigate computational burden. A unified implementation is introduced to avoid the need to rebuild separate synthesis models for varying source-target modality configurations. Comprehensive demonstrations are performed for synthesizing missing sequences in multi-contrast MRI, and CT images from MRI. Our results indicate superiority of ResViT against competing CNN- and transformer-based methods in terms of qualitative observations and quantitative metrics.

Application and prospect for generative adversarial networks in cross-modality reconstruction of medical images

Cross-modality reconstruction of medical images refers to predicting the image from one modality to another so as to achieve more accurate personalized medicine. Generative adversarial networks is the most commonly used deep learning technique in cross-modality reconstruction. It can generate realistic images by learning features from implicit distributions that follow the distributions of real data and then reconstruct the image of another modality rapidly. With the sharp increase in clinical demand for multi-modality medical image, this technology has been widely used in the task of cross modal reconstruction between different medical image modalities, such as magnetic resonance imaging, computed tomography and positron emission computed tomography. It can achieve accurate and efficient cross-modality image reconstruction in different parts of the body, such as the brain, heart, etc. In addition, although GAN has achieved some success in cross-modality reconstruction, its stability, generalization ability, and accuracy still need further research and improvement.

Multimodal image synthesis based on disentanglement representations of anatomical and modality specific features, learned using uncooperative relativistic GAN

Growing number of methods for attenuation-coefficient map estimation from magnetic resonance (MR) images have recently been proposed because of the increasing interest in MR-guided radiotherapy and the introduction of positron emission tomography (PET) MR hybrid systems. We propose a deep-network ensemble incorporating stochastic-binary-anatomical encoders and imaging-modality variational autoencoders, to disentangle image-latent spaces into a space of modality-invariant anatomical features and spaces of modality attributes. The ensemble integrates modality-modulated decoders to normalize features and image intensities based on imaging modality. Besides promoting disentanglement, the architecture fosters uncooperative learning, offering ability to maintain anatomical structure in a cross-modality reconstruction. Introduction of a modality-invariant structural consistency constraint further enforces faithful embedding of anatomy. To improve training stability and fidelity of synthesized modalities, the ensemble is trained in a relativistic generative adversarial framework incorporating multiscale discriminators. Analyses of priors and network architectures as well as performance validation were performed on computed tomography (CT) and MR pelvis datasets. The proposed method demonstrated robustness against intensity inhomogeneity, improved tissue-class differentiation, and offered synthetic CT in Hounsfield units with intensities consistent and smooth across slices compared to the state-of-the-art approaches, offering median normalized mutual information of 1.28, normalized cross correlation of 0.97, and gradient cross correlation of 0.59 over 324 images.

Building and evaluating an artificial intelligence algorithm: A practical guide for practicing oncologists

The use of machine learning and deep learning has enabled many applications, previously thought of as being impossible. Among all medical fields, cancer care is arguably the most significantly impacted, with precision medicine now truly being a possibility. The effect of these technologies, loosely known as artificial intelligence, is particularly striking in fields involving images (such as radiology and pathology) and fields involving large amounts of data (such as genomics). Practicing oncologists are often confronted with new technologies claiming to predict response to therapy or predict the genomic make-up of patients. Underst-anding these new claims and technologies requires a deep understanding of the field. In this review, we provide an overview of the basis of deep learning. We describe various common tasks and their data requirements so that oncologists could be equipped to start such projects, as well as evaluate algorithms presented to them.

Deep learning based domain adaptation for mitochondria segmentation on EM volumes

BACKGROUND AND OBJECTIVE

Accurate segmentation of electron microscopy (EM) volumes of the brain is essential to characterize neuronal structures at a cell or organelle level. While supervised deep learning methods have led to major breakthroughs in that direction during the past years, they usually require large amounts of annotated data to be trained, and perform poorly on other data acquired under similar experimental and imaging conditions. This is a problem known as domain adaptation, since models that learned from a sample distribution (or source domain) struggle to maintain their performance on samples extracted from a different distribution or target domain. In this work, we address the complex case of deep learning based domain adaptation for mitochondria segmentation across EM datasets from different tissues and species.

METHODS

We present three unsupervised domain adaptation strategies to improve mitochondria segmentation in the target domain based on (1) state-of-the-art style transfer between images of both domains; (2) self-supervised learning to pre-train a model using unlabeled source and target images, and then fine-tune it only with the source labels; and (3) multi-task neural network architectures trained end-to-end with both labeled and unlabeled images. Additionally, to ensure good generalization in our models, we propose a new training stopping criterion based on morphological priors obtained exclusively in the source domain. The code and its documentation are publicly available at https://github.com/danifranco/EM_domain_adaptation.

RESULTS

We carried out all possible cross-dataset experiments using three publicly available EM datasets. We evaluated our proposed strategies and those of others based on the mitochondria semantic labels predicted on the target datasets.

CONCLUSIONS

The methods introduced here outperform the baseline methods and compare favorably to the state of the art. In the absence of validation labels, monitoring our proposed morphology-based metric is an intuitive and effective way to stop the training process and select in average optimal models.

Cross-Modality Image Adaptation Based on Volumetric Intensity Gaussian Process Models (VIGPM)

Image-based diagnosis routinely depends on more that one image modality for exploiting the complementary information they provide. However, it is not always possible to obtain images from a secondary modality for several reasons such as cost, degree of invasiveness and non-availability of scanners. Three-dimensional (3D) morphable models have made a significant contribution to the field of medical imaging for feature-based analysis. Here we extend their use to encode 3D volumetric imaging modalities. Specifically, we build a Gaussian Process (GP) over transformations establishing anatomical correspondence between training images within a modality. Given, two different modalities, the GP's eigenspace (latent space) can then be used to provide a parametric representation of each image modality, and we provide an operator for cross-domain translation between the two. We show that the latent space yields samples that are representative of the encoded modality. We also demonstrate that a 3D volumetric image can be efficiently encoded in latent space and transferred to synthesize the corresponding image in another modality. The framework called VIGPM can be extended by designing a fitting process to learn an observation in a given modality and performing cross-modality synthesis. Clinical Relevance- The proposed method provides a way to access a multi modality image from one modality. Both the source and synthetic modalities are in anatomical correspondence giving access to registered complementary information.

Robust-Deep: A Method for Increasing Brain Imaging Datasets to Improve Deep Learning Models' Performance and Robustness

A small dataset commonly affects generalization, robustness, and overall performance of deep neural networks (DNNs) in medical imaging research. Since gathering large clinical databases is always difficult, we proposed an analytical method for producing a large realistic/diverse dataset. Clinical brain PET/CT/MR images including full-dose (FD), low-dose (LD) corresponding to only 5 % of events acquired in the FD scan, non-attenuated correction (NAC) and CT-based measured attenuation correction (MAC) PET images, CT images and T1 and T2 MR sequences of 35 patients were included. All images were registered to the Montreal Neurological Institute (MNI) template. Laplacian blending was used to make a natural presentation using information in the frequency domain of images from two separate patients, as well as the blending mask. This classical technique from the computer vision and image processing communities is still widely used and unlike modern DNNs, does not require the availability of training data. A modified ResNet DNN was implemented to evaluate four image-to-image translation tasks, including LD to FD, LD+MR to FD, NAC to MAC, and MRI to CT, with and without using the synthesized images. Quantitative analysis using established metrics, including the peak signal-to-noise ratio (PSNR), structural similarity index metric (SSIM), and joint histogram analysis was performed for quantitative evaluation. The quantitative comparison between the registered small dataset containing 35 patients and the large dataset containing 350 synthesized plus 35 real dataset demonstrated improvement of the RMSE and SSIM by 29% and 8% for LD to FD, 40% and 7% for LD+MRI to FD, 16% and 8% for NAC to MAC, and 24% and 11% for MRI to CT mapping task, respectively. The qualitative/quantitative analysis demonstrated that the proposed model improved the performance of all four DNN models through producing images of higher quality and lower quantitative bias and variance compared to reference images.

Synthesis of magnetic resonance images from computed tomography data using convolutional neural network with contextual loss function

BACKGROUND

Magnetic resonance imaging (MRI) images synthesized from computed tomography (CT) data can provide more detailed information on pathological structures than that of CT data alone; thus, the synthesis of MRI has received increased attention especially in medical scenarios where only CT images are available. A novel convolutional neural network (CNN) combined with a contextual loss function was proposed for synthesis of T1- and T2-weighted images (T1WI and T2WI) from CT data.

METHODS

A total of 5,053 and 5,081 slices of T1WI and T2WI, respectively were selected for the dataset of CT and MRI image pairs. Affine registration, image denoising, and contrast enhancement were done on the aforementioned multi-modality medical image dataset comprising T1WI, T2WI, and CT images of the brain. A deep CNN was then proposed by modifying the ResNet structure to constitute the encoder and decoder of U-Net, called double ResNet-U-Net (DRUNet). Three different loss functions were utilized to optimize the parameters of the proposed models: mean squared error (MSE) loss, binary crossentropy (BCE) loss, and contextual loss. Statistical analysis of the independent-sample t-test was conducted by comparing DRUNets with different loss functions and different network layers.

RESULTS

DRUNet-101 with contextual loss yielded higher values of peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Tenengrad function (i.e., 34.25±2.06, 0.97±0.03, and 17.03±2.75 for T1WI and 33.50±1.08, 0.98±0.05, and 19.76±3.54 for T2WI respectively). The results were statistically significant at P<0.001 with a narrow confidence interval of difference, indicating the superiority of DRUNet-101 with contextual loss. In addition, both image zooming and difference maps presented for the final synthetic MR images visually reflected the robustness of DRUNet-101 with contextual loss. The visualization of convolution filters and feature maps showed that the proposed model can generate synthetic MR images with high-frequency information.

CONCLUSIONS

The results demonstrated that DRUNet-101 with contextual loss function provided better high-frequency information in synthetic MR images compared with the other two functions. The proposed DRUNet model has a distinct advantage over previous models in terms of PSNR, SSIM, and Tenengrad score. Overall, DRUNet-101 with contextual loss is recommended for synthesizing MR images from CT scans.

Multi-Conditional Constraint Generative Adversarial Network-Based MR Imaging from CT Scan Data

Magnetic resonance (MR) imaging is an important computer-aided diagnosis technique with rich pathological information. The factor of physical and physiological constraint seriously affects the applicability of that technique. Thus, computed tomography (CT)-based radiotherapy is more popular on account of its imaging rapidity and environmental simplicity. Therefore, it is of great theoretical and practical significance to design a method that can construct an MR image from the corresponding CT image. In this paper, we treat MR imaging as a machine vision problem and propose a multi-conditional constraint generative adversarial network (GAN) for MR imaging from CT scan data. Considering reversibility of GAN, both generator and reverse generator are designed for MR and CT imaging, respectively, which can constrain each other and improve consistency between features of CT and MR images. In addition, we innovatively treat the real and generated MR image discrimination as object re-identification; cosine error fusing with original GAN loss is designed to enhance verisimilitude and textural features of the MR image. The experimental results with the challenging public CT-MR image dataset show distinct performance improvement over other GANs utilized in medical imaging and demonstrate the effect of our method for medical image modal transformation.

Synthetic MR: Physical principles, clinical implementation, and new developments

Current clinical MR imaging practices rely on the qualitative assessment of images for diagnosis and treatment planning. While contrast in MR images is dependent on the spin parameters of the imaged tissue, pixel values on MR images are relative and are not scaled to represent any tissue properties. Synthetic MR is a fully featured imaging workflow consisting of efficient multiparameter mapping acquisition, synthetic image generation, and volume quantitation of brain tissues. As the application becomes more widely available on multiple vendors and scanner platforms, it has also gained widespread adoption as clinicians begin to recognize the benefits of rapid quantitation. This review will provide details about the sequence with a focus on the physical principles behind its relaxometry mechanisms. It will present an overview of the products in their current form and some potential issues when implementing it in the clinic. It will conclude by highlighting some recent advances of the technique, including a 3D mapping method and its associated applications. This article is protected by copyright. All rights reserved.

Generative Adversarial Networks in Brain Imaging: A Narrative Review

Artificial intelligence (AI) is expected to have a major effect on radiology as it demonstrated remarkable progress in many clinical tasks, mostly regarding the detection, segmentation, classification, monitoring, and prediction of diseases. Generative Adversarial Networks have been proposed as one of the most exciting applications of deep learning in radiology. GANs are a new approach to deep learning that leverages adversarial learning to tackle a wide array of computer vision challenges. Brain radiology was one of the first fields where GANs found their application. In neuroradiology, indeed, GANs open unexplored scenarios, allowing new processes such as image-to-image and cross-modality synthesis, image reconstruction, image segmentation, image synthesis, data augmentation, disease progression models, and brain decoding. In this narrative review, we will provide an introduction to GANs in brain imaging, discussing the clinical potential of GANs, future clinical applications, as well as pitfalls that radiologists should be aware of.

A Hybrid Deep Learning Method for Early and Late Mild Cognitive Impairment Diagnosis With Incomplete Multimodal Data

Multimodality neuroimages have been widely applied to diagnose mild cognitive impairment (MCI). However, the missing data problem is unavoidable. Most previously developed methods first train a generative adversarial network (GAN) to synthesize missing data and then train a classification network with the completed data. These methods independently train two networks with no information communication. Thus, the resulting GAN cannot focus on the crucial regions that are helpful for classification. To overcome this issue, we propose a hybrid deep learning method. First, a classification network is pretrained with paired MRI and PET images. Afterward, we use the pretrained classification network to guide a GAN by focusing on the features that are helpful for classification. Finally, we synthesize the missing PET images and use them with real MR images to fine-tune the classification model to make it better adapt to the synthesized images. We evaluate our proposed method on the ADNI dataset, and the results show that our method improves the accuracies obtained on the validation and testing sets by 3.84 and 5.82%, respectively. Moreover, our method increases the accuracies for the validation and testing sets by 7.7 and 9.09%, respectively, when we synthesize the missing PET images via our method. An ablation experiment shows that the last two stages are essential for our method. We also compare our method with other state-of-the-art methods, and our method achieves better classification performance.