BACKGROUND:Deep learning (DL)-based image enhancement is widely used to improve suboptimal medical imaging. Whether it also benefits diagnostic-quality MRI in downstream task performance and data-efficiency remains unclear. PURPOSE:To investigate the impact of DL-based enhancement applied to diagnostic quality structural MRI for Alzheimer's disease (AD) classification. STUDY TYPE:Retrospective. POPULATION:A total of 2293 brain MRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) were split into training (n = 1605), validation (n = 229), and internal test (n = 459) sets. Two hundred and seventy scans from the National Alzheimer's Coordinating Center (NACC) served as an external test set. FIELD STRENGTH/SEQUENCE:1.5 T; 3D T1-weighted gradient-echo. ASSESSMENT:Each scan was enhanced by SubtleHD (SHD), an FDA-cleared DL-based MR enhancement tool. ResNet34 and DenseNet121 were trained on standard-of-care (SOC) and SHD-enhanced images to classify subjects as cognitively normal, mild cognitive impairment, or AD and evaluated by accuracy and macro-area under the receiver operating characteristic curve (macro-AUC). Data efficiency was assessed by retraining on stratified training subsets (50%-100%). STATISTICAL TESTS:McNemar test for accuracy and DeLong test for macro-AUC in three-class one-versus-rest setting (p < 0.05). RESULTS:SHD enhancement increased ResNet34 accuracy from 85.2% to 88.7% and macro-AUC from 0.951 to 0.968 (both significant), and DenseNet121 accuracy from 90.2% to 92.2% (p = 0.18) and macro-AUC from 0.978 to 0.982 (p = 0.29). Models trained on 70% of SHD-enhanced dataset matched those trained on the full SOC dataset (accuracy: 85.9%, macro-AUC: 0.942), indicating improved data efficiency with enhancement. In NACC, the SOC-trained model achieved accuracy of 49.2% and macro-AUC of 0.679 versus 63.0% and 0.819 for the SHD-trained model (both significant); the SHD-trained model retained an advantage on unenhanced NACC images (macro-AUC: 0.772). DATA CONCLUSION:DL-based enhancement of diagnostic-quality MRI improves downstream Alzheimer's disease classification performance and reduces the amount of training data required. This suggests that conventional definitions of image quality may underestimate the information content available for machine learning. EVIDENCE LEVEL:3. TECHNICAL EFFICACY STAGE:2.

Perceptual losses are widely used in medical image synthesis because they encourage agreement in high-level structure beyond voxel-wise intensity similarity. In practice, most perceptual losses are still computed with natural-image backbones such as VGG16 or ResNet50, even when the target domain is magnetic resonance imaging (MRI). This mismatch may weaken supervision for anatomy, contrast enhancement, and acquisition variability. We test whether medical foundation model features provide a more suitable perceptual loss for brain MRI contrast dose simulation. The study has two stages. First, we compare RadImageNet, SegVol, and BrainIAC with ImageNet-pretrained VGG16 and ResNet50 as frozen feature extractors on four public medical imaging benchmarks: thyroid ultrasound, breast ultrasound, anterior cruciate ligament knee MRI, and meniscus knee MRI. RadImageNet achieves the lowest mean rank across the Stage I representation suite and is selected as ϕ^⋆. Second, we replace only the VGG16 feature extractor in an existing iterative brain MRI dose simulation framework with ϕ^⋆. The generator, reconstruction loss, adversarial loss, auxiliary losses, optimization schedule, and loss weights are kept unchanged. Standard metrics change modestly, with PSNR increasing from 41.63 to 41.74, SSIM from 0.9739 to 0.9754, RMSE decreasing from 0.1384 to 0.1369, and residual-uptake CNR from 0.0085 to 0.0082. The visual results show the main effect: RadImageNet reduces residual enhancement in marked structures, follows a more faithful dose-reduction trajectory, and remains close to the acquired 10

Positron emission tomography (PET) with [F-18]-fludeoxyglucose (FDG) can visualize the spatial pattern of neurodegeneration-related glucose hypometabolism. We proposed the "MRI-styled PET," leveraging anatomical information from T1-weighted magnetic resonance imaging to enhance the structural details and quantitative accuracy of FDG-PET, which is degraded by partial volume effects (PVE). The proposed framework comprised a baseline encoder-decoder image fusion model and several task-specific modules; notably, the alternative anatomical input significantly contributes to correcting the under/overestimation of gray/white matter while the adaptive multiscale structural similarity loss utilized learnable ratios across various receptive fields to modulate attention to tissue contrast. Compared to a traditional anatomy-guided post-reconstruction PVE correction method (PVC-PET), MRI-styled PET demonstrated significantly higher structural similarity and peak signal-to-noise ratio than the baseline image fusion model (Baseline), showcasing the effectiveness of the proposed task-specific modules. In several Alzheimer's Disease-related brain regions, MRI-styled PET exhibited consistent increases in corrective effects regardless of disease stage, compared to Baseline and PVC-PET. In conclusion, this study represented an initial exploration of a deep-learning approach for correcting PVE in PET without prior knowledge regarding the correction method or the underlying radiotracer uptake and without assumptions about the system point-spread function. Our implementation is available at https://github.com/NTUMMIO/MRI-styled-PET.

Multi-contrast magnetic resonance imaging (MRI) is widely used in clinical practice as each contrast provides complementary information. However, the availability of each imaging contrast may vary amongst patients, which poses challenges to radiologists and automated image analysis algorithms. A general approach for tackling this problem is missing data imputation, which aims to synthesize the missing contrasts from existing ones. While several convolutional neural networks (CNN) based algorithms have been proposed, they suffer from the fundamental limitations of CNN models, such as the requirement for fixed numbers of input and output channels, the inability to capture long-range dependencies, and the lack of interpretability. In this work, we formulate missing data imputation as a sequence-to-sequence learning problem and propose a multi-contrast multi-scale Transformer (MMT), which can take any subset of input contrasts and synthesize those that are missing. MMT consists of a multi-scale Transformer encoder that builds hierarchical representations of inputs combined with a multi-scale Transformer decoder that generates the outputs in a coarse-to-fine fashion. The proposed multi-contrast Swin Transformer blocks can efficiently capture intra- and inter-contrast dependencies for accurate image synthesis. Moreover, MMT is inherently interpretable as it allows us to understand the importance of each input contrast in different regions by analyzing the in-built attention maps of Transformer blocks in the decoder. Extensive experiments on two large-scale multi-contrast MRI datasets demonstrate that MMT outperforms the state-of-the-art methods quantitatively and qualitatively.

This study introduces a novel image modality, named MRI-styled PET, which aims to enhance local structural details in FDG-PET images using T1-weighted MRI, with minimal distortion in SUVr. The proposed model involves a UNet++ network with a fusion strategy block in the bottleneck layer to combine multi-scale features from PET and MRI. Three fusion strategies, namely spatial-channel attention, self-attention, and cross-attention, have been proposed. The effectiveness of the fusion models was evaluated based on quantitative metrics, regional SUVr consistency, and visual realism. The results showed that spatial-channel attention with only the gray matter portion of MRI as input is the optimal fusion model. These preliminary results demonstrate the potential of MRI-styled PET images as a PET enhancement method with clinical utility.

Complementary information from multi-contrast MRI data is used in deep learning algorithms for reducing contrast dosage in brain MRI. Though existing models produce clinically equivalent post-contrast images, they lack explainability in terms of mapping the source of contrast information from input to output. In this work we explore the feasibility of an explainable deep learning model for gadolinium dose reduction in contrast-enhanced brain MRI.

With the aim of improving the performance of an automated quality control system, we propose to use a data augmentation technique based on cropped patches of simulated artifacts (CutArt) instead of artifacts that are distributed across the entire image. This has the advantage of improving the artifact localization and quality control classification performance, as assessed by experiments on simulated as well as real artifact affected data. Localization experiments suggested that the CutArt model learns to focus on the tissue of interest instead of the image background.

Purpose: Perfusion imaging assesses target mismatch but requires contrast and processing software. Clinical/diffusion mismatch can miss cases that have target mismatch and could benefit from thrombectomy. We explored whether a neural network can predict hypoperfusion and identify target mismatch from diffusion-weighted imaging (DWI) and clinical information alone. Methods: Acute ischemic stroke cases with baseline MR perfusion and DWI were included from two multi-center trials and one registry for model development and a separate randomized trial for external validation. MR perfusion images were processed by RAPID, which segments Tmax lesion (Tmax≥6s) and the ischemic core lesion (apparent diffusion coefficient [ADC]≤ 620). A 3D U-Net was trained using baseline DWI, ADC, NIH stroke scale, and side of stroke as input, and the union of Tmax and ischemic core segmentation as the ground truth. 5-fold cross-validation was performed for model development cohort. Model performance was evaluated by Dice score coefficient (DSC) and volume difference. Sensitivity and specificity of model target mismatch and clinical/diffusion mismatch criteria from the DAWN were compared, using the DEFUSE 3 target mismatch as reference. Results: 413 patients were included for model development and 46 for external validation. In model development and external validation cohort, the model achieved median DSC of 0.61 (IQR 0.45, 0.71) and 0.62 (IQR 0.53, 0.72); and volume difference of 3 ml (IQR -37, 41) and 7 ml (IQR -24, 32), respectively. Compared to the clinical/diffusion mismatch approach, the model identified target mismatch with a sensitivity of 89.5% vs 49.3%, a specificity of 77.5% vs 89.2% in the model development cohort, and a sensitivity of 95.6% vs 41.3% in external validation cohort. Conclusion: A 3D U-Net can predict hypoperfusion lesions from baseline DWI and clinical information, with more sensitive classification of target mismatch than clinical/diffusion mismatch.

Artificial intelligence (AI)-based image enhancement has the potential to reduce scan times while improving signal-to-noise ratio (SNR) and maintaining spatial resolution. This study prospectively evaluated AI-based image enhancement in 32 consecutive patients undergoing clinical brain MRI. Standard-of-care (SOC) three-dimensional (3D) T1 precontrast, 3D T2 fluid-attenuated inversion recovery, and 3D T1 postcontrast sequences were performed along with 45% faster versions of these sequences using half the number of phase-encoding steps. Images from the faster sequences were processed by a Food and Drug Administration-cleared AI-based image enhancement software for resolution enhancement. Four board-certified neuroradiologists scored the SOC and AI-enhanced image series independently on a five-point Likert scale for image SNR, anatomic conspicuity, overall image quality, imaging artifacts, and diagnostic confidence. While interrater κ was low to fair, the AI-enhanced scans were noninferior for all metrics and actually demonstrated a qualitative SNR improvement. Quantitative analyses showed that the AI software restored the high spatial resolution of small structures, such as the septum pellucidum. In conclusion, AI-based software can achieve noninferior image quality for 3D brain MRI sequences with a 45% scan time reduction, potentially improving the patient experience and scanner efficiency without sacrificing diagnostic quality. Keywords: MR Imaging, CNS, Brain/Brain Stem, Reconstruction Algorithms © RSNA, 2022.

PurposeWith rising safety concerns over the use of gadolinium‐based contrast agents (GBCAs) in contrast‐enhanced MRI, there is a need for dose reduction while maintaining diagnostic capability. This work proposes comprehensive technical solutions for a deep learning (DL) model that predicts contrast‐enhanced images of the brain with approximately 10% of the standard dose, across different sites and scanners.MethodsThe proposed DL model consists of a set of methods that improve the model robustness and generalizability. The steps include multi‐planar reconstruction, 2.5D model, enhancement‐weighted L1, perceptual, and adversarial losses. The proposed model predicts contrast‐enhanced images from corresponding pre‐contrast and low‐dose images. With IRB approval and informed consent, 640 heterogeneous patient scans (56 train, 13 validation, and 571 test) from 3 institutions consisting of 3D T1‐weighted brain images were used. Quantitative metrics were computed and 50 randomly sampled test cases were evaluated by 2 board‐certified radiologists. Quantitative tumor segmentation was performed on cases with abnormal enhancements. Ablation study was performed for systematic evaluation of proposed technical solutions.ResultsThe average peak signal‐to‐noise ratio (PSNR) and structural similarity index measure (SSIM) between full‐dose and model prediction were dB and , respectively. Radiologists found the same enhancing pattern in 45/50 (90%) cases; discrepancies were minor differences in contrast intensity and artifacts, with no effect on diagnosis. The average segmentation Dice score between full‐dose and synthesized images was (median = 0.91).ConclusionsWe have proposed a DL model with technical solutions for low‐dose contrast‐enhanced brain MRI with potential generalizability under diverse clinical settings.

Background: Magnetic resonance (MR) scans are routine clinical procedures for monitoring people with multiple sclerosis (PwMS). Patient discomfort, timely scheduling, and financial burden motivate the need to accelerate MR scan time. We examined the clinical application of a deep learning (DL) model in restoring the image quality of accelerated routine clinical brain MR scans for PwMS. Methods: We acquired fast 3D T1w BRAVO and fast 3D T2w FLAIR MRI sequences (half the phase encodes and half the number of slices) in parallel to conventional parameters. Using a subset of the scans, we trained a DL model to generate images from fast scans with quality similar to the conventional scans and then applied the model to the remaining scans. We calculated clinically relevant T1w volumetrics (normalized whole brain, thalamic, gray matter, and white matter volume) for all scans and T2 lesion volume in a sub-analysis. We performed paired t-tests comparing conventional, fast, and fast with DL for these volumetrics, and fit repeated measures mixed-effects models to test for differences in correlations between volumetrics and clinically relevant patient-reported outcomes (PRO). Results: We found statistically significant but small differences between conventional and fast scans with DL for all T1w volumetrics. There was no difference in the extent to which the key T1w volumetrics correlated with clinically relevant PROs of MS symptom burden and neurological disability. Conclusion: A deep learning model that improves the image quality of the accelerated routine clinical brain MR scans has the potential to inform clinically relevant outcomes in MS.

Fluoroscopic imaging relies on ionizing radiation to provide physicians with high quality video feedback during a surgical operation. Radiation exposure is harmful for both the physician and patient, but reducing dosage results in a much noisier video. We hence propose an algorithm that delivers the same quality video with 4× reduction in radiation dose. Our method is a deep learning approximation to VBM4D, a state-of-the-art video denoiser. Neither VBM4D nor previous deep learning methods are clinically feasible, however, as their high inference runtimes prohibit live display on an operating room monitor. On the other hand, we present a video denoising method which executes orders of magnitude faster while achieving state-of-the-art performance. This provides compelling potential for real-time clinical application in fluoroscopic imaging.

Positron emission tomography (PET) is a widely used molecular imaging modality for various clinical applications. With Magnetic Resonance Imaging (MRI) providing anatomical information, simultaneous PET/MR reduces the radiation risk. Both improved hardware and algorithms have been developed to further reduce the amount of radiotracer dosage, but these methods are not yet applied to very low dose. Here, we propose a Deep Learning based method to enable ultra-low-dose PET denoising with multi-contrast information from simultaneous MRI. Methods:The method is implemented to denoise 18F-fluorodeoxyglucose (FDG) brain PET images from low-dose images with 200-fold dose reduction through undersampling, and evaluated for glioblastoma (GBM) patients. Comprehensive quantitative and qualitative evaluations were conducted to verify the performance and clinical applicability of the proposed method, including quantitative accuracy evaluation, visual quality evaluation, reader study with manual tumor segmentation to evaluate the diagnostic quality. Results:The results demonstrate that the proposed method achieves superior results in performance and efficiency comparing with the state-of-art denoising methods. Conclusion:Though reconstructed from scans with only 0.5% of the standard dose, the denoised ultra-low-dose PET images deliver similar visual quality and diagnostic information as the standard-dose PET images. By combining PET and MR information, the proposed Deep Learning based method improves image quality of ultra-low-dose PET, preserves diagnostic quality, and potentially enables much safer, faster, and more cost-effective PET/MR studies.

Multi-contrast Magnetic Resonance Imaging (MRI) acquisitions from a single scan have tremendous potential to streamline exams and reduce imaging time. However, maintaining clinically feasible scan time necessitates significant undersampling, pushing the limits on compressed sensing and other low-dimensional techniques. During MRI scanning, one of the possible solutions is by using undersampling designs which can effectively improve the acquisition and achieve higher reconstruction accuracy. However, existing undersampling optimization methods are time-consuming and the limited performance prevents their clinical applications. In this paper, we proposed an improved undersampling trajectory optimization scheme to generate an optimized trajectory within seconds and apply it to subsequent multi-contrast MRI datasets on a per-subject basis, where we named it OUTCOMES. By using a data-driven method combined with improved algorithm design, GPU acceleration, and more efficient computation, the proposed method can optimize a trajectory within 5-10 seconds and achieve 30%-50% reconstruction improvement with the same acquisition cost, which makes real-time under-sampling optimization possible for clinical applications.

Arterial Spin Labeling (ASL) is a popular non-invasive neuroimaging technique to use MRI for quantitatively Cerebral Blood Flow (CBF) mapping. However, ASL usually suffers from poor signal quality and repeated measurements are typically acquired to improve signal quality through averaging at the cost of long scan time. In this work, a deep learning algorithm is proposed to leverage both convolutional neural network (CNN) based image enhancement as well as combining complementary/mutual information from multiple tissue contrasts in ASL acquisition. Both quantitative and qualitative evaluation demonstrate the performance and stability of the proposed algorithm and its superiority over conventional denoising algorithms and standard deep learning based denoising. The results demonstrate the feasibility of efficient and high-quality ASL measurements from average-free fast acquisition which will enable broader clinical application of ASL.

Objective: Ischemic core prediction from CT perfusion (CTP) remains inaccurate compared with gold standard diffusion-weighted imaging (DWI). We evaluated if a deep learning model to predict the DWI lesion from MR perfusion (MRP) could facilitate ischemic core prediction on CTP. Method: Using the multi-center CRISP cohort of acute ischemic stroke patient with CTP before thrombectomy, we included patients with major reperfusion (TICI score≥2b), adequate image quality, and follow-up MRI at 3-7 days. Perfusion parameters including Tmax, mean transient time, cerebral blood flow (CBF), and cerebral blood volume were reconstructed by RAPID software. Core lab experts outlined the stroke lesion on the follow-up MRI. A previously trained MRI model in a separate group of patients was used as a starting point, which used MRP parameters as input and RAPID ischemic core on DWI as ground truth. We fine-tuned this model, using CTP parameters as input, and follow-up MRI as ground truth. Another model was also trained from scratch with only CTP data. 5-fold cross validation was used. Performance of the models was compared with ischemic core (rCBF≤30%) from RAPID software to identify the presence of a large infarct (volume>70 or >100ml). Results: 94 patients in the CRISP trial met the inclusion criteria (mean age 67±15 years, 52% male, median baseline NIHSS 18, median 90-day mRS 2). Without fine-tuning, the MRI model had an agreement of 73% in infarct >70ml, and 69% in >100ml; the MRI model fine-tuned on CT improved the agreement to 77% and 73%; The CT model trained from scratch had agreements of 73% and 71%; All of the deep learning models outperformed the rCBF segmentation from RAPID, which had agreements of 51% and 64%. See Table and figure. Conclusions: It is feasible to apply MRP-based deep learning model to CT. Fine-tuning with CTP data further improves the predictions. All deep learning models predict the stroke lesion after major recanalization better than thresholding approaches based on rCBF.

Deep learning models with fine-tuning lead to better performance for predicting tissue at risk and ischemic core, outperforming conventional thresholding methods. BACKGROUND AND PURPOSE: In acute stroke patients with large vessel occlusions, it would be helpful to be able to predict the difference in the size and location of the final infarct based on the outcome of reperfusion therapy. Our aim was to demonstrate the value of deep learning–based tissue at risk and ischemic core estimation. We trained deep learning models using a baseline MR image in 3 multicenter trials. MATERIALS AND METHODS: Patients with acute ischemic stroke from 3 multicenter trials were identified and grouped into minimal (≤20%), partial (20%-80%), and major (≥80%) reperfusion status based on 4- to 24-hour follow-up MR imaging if available or into unknown status if not. Attention-gated convolutional neural networks were trained with admission imaging as input and the final infarct as ground truth. We explored 3 approaches: 1) separate: train 2 independent models with patients with minimal and major reperfusion; 2) pretraining: develop a single model using patients with partial and unknown reperfusion, then fine-tune it to create 2 separate models for minimal and major reperfusion; and 3) thresholding: use the current clinical method relying on apparent diffusion coefficient and time-to-maximum of the residue function maps. Models were evaluated using area under the curve, the Dice score coefficient, and lesion volume difference. RESULTS: Two hundred thirty-seven patients were included (minimal, major, partial, and unknown reperfusion: n = 52, 80, 57, and 48, respectively). The pretraining approach achieved the highest median Dice score coefficient (tissue at risk = 0.60, interquartile range, 0.43–0.70; core = 0.57, interquartile range, 0.30–0.69). This was higher than the separate approach (tissue at risk = 0.55; interquartile range, 0.41–0.69; P = .01; core = 0.49; interquartile range, 0.35–0.66; P = .04) or thresholding (tissue at risk = 0.56; interquartile range, 0.42–0.65; P = .008; core = 0.46; interquartile range, 0.16–0.54; P < .001). CONCLUSIONS: Deep learning models with fine-tuning lead to better performance for predicting tissue at risk and ischemic core, outperforming conventional thresholding methods.

Objective: We previously developed two separate deep learning (DL) models to segment the ischemic core and critically hypoperfused tissue on baseline imaging of acute ischemic stroke patients. We aimed to validate the models in an external, multi-center randomized clinical trial (DEFUSE3) and compare with the current clinical standard. Methods: The DL models were previously trained in a separate dataset in which follow-up MRI, obtained at 3-7 days, was used as the reference for critically hypoperfused tissue in patients who did not reperfuse and as the reference for the ischemic core in patients who did reperfuse. For validation, we included DEFUSE3 patients with adequate quality baseline MR perfusion and a 24-hour follow-up DWI scan. The 24-hour DWI lesion served as the reference for ischemic core in patients in the thrombectomy arm and for critically hypoperfused tissue for patients in the medical arm. RAPID was used to generate perfusion maps (Tmax, cerebral blood flow, cerebral blood volume, and mean transient time). The accuracy of segmenting the ischemic core and critically hypoperfused tissue on baseline imaging was compared between the DL approach and the traditional thresholding approach implemented in RAPID. Results: In the 46 patients included for analysis, 24 were in the medical arm and 22 in the thrombectomy arm. Compared to a traditional thresholding method, the DL model segmented the ischemic core more accurately (AUC of 0.92 vs 0.72, p=0.0001 and volume difference of -8ml vs -21ml, p=0.001). Similarly, the DL model segmented critically hypoperfused tissue more accurately (AUC of 0.93 vs 0.80, p<0.0001; volume difference 14ml vs. 55ml, p=0.0005). However, great heterogeneity in final infarct was noticed in medical arm. See tables and figures. Conclusions: The DL-based critical hypoperfusion and ischemic core prediction provides more accurate prediction on final infarct than a commonly used thresholding method in this external validation.

More widespread use of positron emission tomography (PET) imaging is limited by its high cost and radiation dose. Reductions in PET scan time or radiotracer dosage typically degrade diagnostic image quality (DIQ). Deep-learning-based reconstruction may improve DIQ, but such methods have not been clinically evaluated in a realistic multicenter, multivendor environment. In this study, we evaluated the performance and generalizability of a deep-learning-based image-quality enhancement algorithm applied to fourfold reduced-count whole-body PET in a realistic clinical oncologic imaging environment with multiple blinded readers, institutions, and scanner types. We demonstrate that the low-count-enhanced scans were noninferior to the standard scans in DIQ ( p < 0.05) and overall diagnostic confidence ( p < 0.001) independent of the underlying PET scanner used. Lesion detection for the low-count-enhanced scans had a high patient-level sensitivity of 0.94 (0.83–0.99) and specificity of 0.98 (0.95–0.99). Interscan kappa agreement of 0.85 was comparable to intrareader (0.88) and pairwise inter-reader agreements (maximum of 0.72). SUV quantification was comparable in the reference regions and lesions (lowest p -value=0.59) and had high correlation (lowest CCC = 0.94). Thus, we demonstrated that deep learning can be used to restore diagnostic image quality and maintain SUV accuracy for fourfold reduced-count PET scans, with interscan variations in lesion depiction, lower than intra- and interreader variations. This method generalized to an external validation set of clinical patients from multiple institutions and scanner types. Overall, this method may enable either dose or exam-duration reduction, increasing safety and lowering the cost of PET imaging.