Purpose: Synthetic MRI can generate multiple contrasts from a single acquisition, yet synthetic fluid-attenuated inversion recovery (FLAIR) generally shows lower quality than conventional FLAIR. We aimed to improve 3D synthetic FLAIR image quality using deep learning, without losing the scan-time advantage of synthetic MRI.

Methods: We studied 55 adults with inflammatory demyelinating diseases who underwent 3T MRI. For each participant, five 3D quantification using an interleaved Look–Locker acquisition sequence with T2 preparation (3D-QALAS) source images and a conventional 3D-FLAIR image were acquired, and a synthetic FLAIR image was generated from the 3D-QALAS data. We trained a deep learning model in which a 3D U-shaped convolutional network (U-Net)–based attention network predicted voxel-wise weights for generating FLAIR images from the five 3D-QALAS source images, using conventional 3D-FLAIR images as the reference. The model was trained with a combined loss function of mean squared error, content loss, and style loss. Agreement with the reference was assessed using image similarity and error metrics, lesion overlap using the Dice similarity coefficient, and overall image quality and focal lesion visibility using blinded radiologist ratings. Synthetic FLAIR and the deep learning–generated FLAIR images were compared using 2-sided Wilcoxon signed-rank tests; P < 0.05 was considered statistically significant.

Results: The deep learning–generated FLAIR images showed significantly higher agreement with the reference image than synthetic FLAIR images (all P < 0.001). Lesion overlap was higher with the deep learning–generated FLAIR images (median [interquartile range]: 0.642 [0.528–0.711] versus 0.487 [0.344–0.641]). Qualitatively, the deep learning–generated FLAIR images improved overall image quality and focal lesion visibility, but reader scores remained lower than those for the reference conventional 3D-FLAIR images (all P < 0.001).

Conclusion: A deep learning–based approach applied to five 3D-QALAS source images improved the image quality of 3D synthetic FLAIR. These improvements may increase the clinical utility of 3D synthetic MRI for neuroradiologic assessment.

Purpose: Most existing multi-contrast MRI super-resolution (MCMSR) methods rely on spatial-domain fusion and global attention, overlooking explicit high-frequency (HF) priors while incurring high computational costs. This work addresses these limitations through a general reference-guided MCMSR framework designed for low computational cost, applicable to any contrast pairing rather than a fixed clinical protocol.

Methods: We introduce a wavelet-guided HF prior modeling block for directional-based decomposition and bounded nonlinear enhancement, enabling precise extraction and controlled amplification of anatomical details from both reference and target contrasts. We further introduce a triple cross-contrast fusion module, based on a windowed cross-contrast attention module, to efficiently transfer high-frequency information between contrasts and reduce computational complexity compared to global attention schemes. Additionally, to reduce feature differences across contrasts, a consistent feature fusion module with selective spatial adaptive modulation is incorporated.

Results: Extensive experiments on IXI and M4Raw datasets demonstrate that our proposed windowed cross-contrast attention network (WCCAN) framework consistently outperforms state-of-the-art single and multi-contrast MRI SR methods in terms of quantitative accuracy and visual fidelity. In addition, the WCCAN model achieves lower computational complexity and faster inference time compared to the other MCMSR methods.

Conclusion: The proposed WCCAN framework provides an efficient and accurate solution for MCMSR, demonstrating superior reconstruction quality with reduced computational cost compared to existing methods.

Purpose: To evaluate the performance of a Transformer-based U-Net, TransDisCo, for distortion correction in clinical diffusion-weighted imaging (DWI), including high b-value DWI, diffusion tensor imaging (DTI), and diffusion kurtosis imaging (DKI) and, furthermore, to compare its efficacy with an established U-Net–based method (Synb0) on a clinical dataset including lesion cases.

Methods: This study utilized a clinical dataset acquired at the University of Tokyo Hospital, comprising DWI data from patients (including cases with brain tumors and cerebrovascular disorders) suitable for DTI and DKI analysis. We compared the distortion-correction performance of our proposed TransDisCo with that of Synb0, a U-Net–based model. Both methods were trained on T1-weighted images and uncorrected DWI images, with FMRIB Software Library–corrected DWI images used as ground truth. To assess the contribution of the self-attention mechanism in TransDisCo, an ablation study was conducted using a convolutional neural network (CNN)–based variant, in which encoder Transformer blocks were replaced with convolutional blocks. We evaluated corrected DWI images and derived DTI/DKI maps using peak SNR, structural similarity, mutual information, and Dice coefficients.

Results: All 3 models corrected distortions in the clinical DWI datasets, including high b-value images and derived DTI/DKI maps. Furthermore, TransDisCo demonstrated significantly superior performance compared to Synb0 across quantitative metrics, particularly in preserving anatomical details and reducing artifacts, even in the presence of pathological changes. Additionally, in the ablation study, TransDisCo outperformed the CNN-based variant trained under identical conditions, indicating that the self-attention mechanism, which enables content-adaptive long-range feature interactions, of the Transformer-based encoder contributed to the observed performance gains.

Conclusion: TransDisCo alleviates limitations of existing CNN-based models and represents a promising approach to DWI distortion correction that surpasses conventional U-Net–based methods on clinical data with diverse pathologies. It has the potential to improve the precision and reliability of diffusion MRI across a wide range of imaging conditions and clinical applications.

Purpose: To develop a fast and precise method for searching rectangular regions in brain tumor images.

Methods: The authors propose a new method for searching rectangular tumor regions in brain MR images. The proposed method consisted of a segmentation network and a fast search method with a user-controllable search metric. As the segmentation network, the U-Net whose encoder was replaced by the EfficientNet was used. In the fast search method, summed-area tables were used for accelerating sums of voxels in rectangular regions. Use of the summed-area tables enabled exhaustive search of the 3D offset (3D full search). The search metric was designed for giving priority to nearly isotropic rectangles over undesirable thin rectangles, and assigning better values for higher tumor fractions even if they exceeded target tumor fractions. The proposed computation and metric were compared with those used in a conventional method using the Brain Tumor Image Segmentation dataset.

Results: When the 3D full search was used, the proposed computation (8 seconds) was 100–500 times faster than the conventional computation (11–40 minutes). When the user-controllable parts of the search metrics were changed variously, the tumor fractions of the proposed metric were higher than those of the conventional metric. In addition, the conventional metric preferred undesirable thin rectangles whereas the proposed metric preferred nearly isotropic rectangles.

Conclusion: The proposed method is promising for implementing fast and precise search of rectangular tumor regions, which is useful for brain tumor diagnosis using MRI systems. The proposed computation reduced processing times of the 3D full search, and the proposed metric improved the quality of the assigned rectangular tumor regions.

Purpose: Glioblastoma (GBM) is an aggressive brain tumor with poor prognosis. O6-methylguanine-DNA-methyltransferase (MGMT) promoter methylation is a critical biomarker for guiding chemotherapy decisions, yet current testing requires invasive tissue sampling. This study aimed to systematically evaluate the diagnostic accuracy of artificial intelligence (AI) models using MRI for non-invasive prediction of MGMT promoter methylation status in GBM.

Methods: We conducted a systematic search of PubMed, ScienceDirect, Scopus, Google Scholar, Cochrane, Web of Science and EMBASE, identifying 480 records. After duplicate removal and screening, 14 studies met inclusion criteria. Data extracted included AI model architecture, MRI sequences, segmentation methods, and diagnostic metrics. A bivariate random-effects model was used to pool sensitivity and specificity. Meta-regression analyses assessed the effect of AI model type on diagnostic performance. Study quality was evaluated using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool.

Results: The bivariate random-effects model yielded a pooled sensitivity of 0.536 (95% confidence interval [95% CI]: 0.509–0.563) and a pooled specificity of 0.514 (95% CI: 0.454–0.574), indicating moderate between-study heterogeneity, with an area under the curve of 0.56. The best-performing models included MGMT-net and transformer-based architectures, particularly when using multimodal MRI inputs. Studies employing automated segmentation and single-sequence input (e.g., T2-weighted only) generally demonstrated lower performance. QUADAS-2 assessment indicated a low risk of bias in most domains, with concerns regarding index test thresholds and external validation in some studies.

Conclusion: AI-based MRI models show moderate-to-high potential for non-invasive MGMT methylation prediction in GBM. However, heterogeneity in study design, imaging protocols, and validation approaches highlights the need for standardized methodologies and robust external validation before clinical adoption.

Purpose: Interpretability and robustness are both critical for developing trustworthy artificial intelligence, especially in high-stakes domains such as medical diagnosis. However, few studies have explored how to enhance robustness within interpretable model frameworks. This work aims to improve the robustness of interpretable multimodal medical imaging diagnostic models, particularly under missing modality conditions.

Methods: We propose the Modality-Robust and Explainable Network (MoRE-Net), a robust and interpretable model for brain tumor grading. Built on a variant of the interpretable prototypical part network, MoRE-Net uses independent per-modality encoders to extract modality-specific features. To address the absence of inter-modality interactions, we introduce 2 key designs: (1) Mamba-based per-modality encoders for efficient global-context modeling; and (2) an online multimodal teacher that guides the per-modality encoders via an alignment loss during early training, which is gradually annealed and removed. We evaluate MoRE-Net on 369 subjects with multimodal MRI from the BraTS2020 dataset, using balanced accuracy (BAC) for grading performance and activation precision (AP) for interpretability. We further validate the model on the real-world ReMIND dataset.

Results: MoRE-Net achieves an average BAC of 73.5% and AP of 61.2% across all missing modality scenarios on BraTS2020 dataset, surpassing baseline methods by about 15% and 21%, respectively. Results on ReMIND dataset and ablation studies confirm its effectiveness of each proposed strategy and the overall robustness.

Conclusion: We introduce MoRE-Net, a novel interpretable and modality-robust model for brain tumor grading. Experimental results demonstrate its strong performance in both diagnostic accuracy and interpretability under missing modality conditions, indicating its potential for clinical deployment.

Purpose: To develop and evaluate short-TR acquisition time-of-flight (STRA-TOF) MR angiography (MRA), which combines an optimized STRA with deep learning–based reconstruction to achieve scan-time reduction while maintaining image quality in the visualization of intracranial arteries.

Methods: Ten healthy volunteers and 3 patients with moyamoya disease were examined using 3D TOF MRA with the clinical moyamoya protocol and 2 STRA-TOF protocols employing 4-slab (STRA4) and 9-slab (STRA9) configurations. STRA-TOF employed a TR of approximately 10 ms with variable-density Poisson-disc sampling and unrolled deep learning reconstruction. Bloch equation simulations validated the theoretical basis for STRA. Quantitative assessment included SNR and contrast-to-noise ratio measurements. Two radiologists independently evaluated image quality using a 3-point scale across 9 vascular territories and overall image quality, with blinded assessment. Statistical analysis was performed using the Friedman test with post hoc Wilcoxon signed-rank tests.

Results: STRA-TOF achieved approximately a 50% reduction in scan time compared with conventional protocols. Both STRA sequences demonstrated significantly higher SNR and contrast-to-noise ratio than conventional TOF (P < 0.001). Overall image quality scores were higher for STRA4 and STRA9 compared with conventional TOF in both readers. Across the 9 vascular territories, both readers consistently rated STRA sequences equal to or superior to conventional TOF, particularly for distal branches. In the small patient cohort with moyamoya disease (n = 3), STRA-TOF demonstrated feasibility for visualizing complex arterial pathology, including stenotic vessels, collateral circulation, and postsurgical vascular changes.

Conclusion: STRA-TOF achieved approximately a 50% reduction in scan time while maintaining or improving image quality compared with conventional 3D TOF MRA. This technique addresses the fundamental scan-time limitations of the conventional method, with potential for significant clinical benefits in terms of patient comfort, workflow efficiency, and improved access to intracranial artery evaluation.

This study evaluated the utility of deep learning reconstruction (DLR) in vessel wall imaging (VWI) for visualizing the entire cerebral arterial system, including cortical arteries. Seventeen patients underwent post-contrast 3D T1WI-CUBE VWI with 0.5 mm isotropic resolution. Images with and without DLR were compared using qualitative and quantitative assessments. Qualitative image quality was rated on a 4-point scale across 29 arterial segments, including the internal carotid, vertebral, basilar, and the 1st to 4th segments of the major cerebral arteries. Quantitative evaluation of the vertebral artery wall assessed SNR and contrast-to-noise ratio (CNR). DLR significantly improved overall image quality compared to the without-DLR group, with cortical arteries rated as optimal in all cases with DLR (all P < 0.001). SNR and CNR were also significantly higher with DLR (P = 0.004). These results suggest that DLR enables high-resolution VWI of intracranial cortical arteries within a clinically acceptable scan time.

Purpose: Among patients with hepatitis B virus-associated liver cirrhosis (HBV-LC), there may be differences in the hepatic parenchyma between those with and without hepatocellular carcinoma (HCC). Proton MR spectroscopy (¹H-MRS) is a well-established tool for noninvasive metabolomics, but has been challenging in the liver allowing only a few metabolites to be detected other than lipids. This study aims to explore the potential of ¹H-MRS of the liver in conjunction with deep learning to differentiate between HBV-LC patients with and without HCC.

Methods: Between August 2018 and March 2021, ¹H-MRS data were collected from 37 HBV-LC patients who underwent MRI for HCC surveillance, without HCC (HBV-LC group, n = 20) and with HCC (HBV-LC-HCC group, n = 17). Based on a priori knowledge from the first 10 patients from each group, big spectral datasets were simulated to develop 2 kinds of convolutional neural networks (CNNs): CNNs quantifying 15 metabolites and 5 lipid resonances (qCNNs) and CNNs classifying patients into HBV-LC and HBV-LC-HCC (cCNNs). The performance of the cCNNs was assessed using the remaining patients in the 2 groups (10 HBV-LC and 7 HBV-LC-HCC patients).

Results: Using a simulated dataset, the quantitative errors with the qCNNs were significantly lower than those with a conventional nonlinear-least-squares-fitting method for all metabolites and lipids (P ≤0.004). The cCNNs exhibited sensitivity, specificity, and accuracy of 100% (7/7), 90% (9/10), and 94% (16/17), respectively, for identifying the HBV-LC-HCC group.

Conclusion: Deep-learning-aided ¹H-MRS with data augmentation by spectral simulation may have potential in differentiating between HBV-LC patients with and without HCC.

This study presents a novel data augmentation approach to improve deep learning (DL)-based segmentation for 3D phase-contrast magnetic resonance angiography (PC-MRA) images affected by pulsation artifacts. Augmentation was achieved by simulating pulsation artifacts through the addition of periodic errors in k-space magnitude. The approach was evaluated on PC-MRA datasets from 16 volunteers, comparing DL segmentation with and without pulsation artifact augmentation to a level-set algorithm. Results demonstrate that DL methods significantly outperform the level-set approach and that pulsation artifact augmentation further improves segmentation accuracy, especially for images with lower velocity encoding. Quantitative analysis using Dice-Sørensen coefficient, Intersection over Union, and Average Symmetric Surface Distance metrics confirms the effectiveness of the proposed method. This technique shows promise for enhancing vascular segmentation in various anatomical regions affected by pulsation artifacts, potentially improving clinical applications of PC-MRA.

Purpose: To evaluate the effect of model-based deep-learning reconstruction (DLR) compared with that of compressed sensing–sensitivity encoding (CS) on cine cardiac magnetic resonance (CMR).

Methods: Cine CMR images of 10 healthy volunteers were obtained with reduction factors of 2, 4, 6, and 8 and reconstructed using CS and DLR. The visual image quality scores assessed sharpness, image noise, and artifacts. Left-ventricular (LV) end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), and ejection fraction (EF) were manually measured. LV global circumferential strain (GCS) was automatically measured using the software. The precision of EDV, ESV, SV, EF, and GCS measurements was compared between CS and DLR using Bland–Altman analysis with full-sampling data as the gold standard.

Results: Compared with CS, DLR significantly improved image quality with reduction factors of 6 and 8. The precision of EDV and ESV with a reduction factor of 8, and GCS with reduction factors of 6 and 8 measurements improved with DLR compared with CS, whereas those of SV and EF measurements were not different between DLR and CS.

Conclusion: The effect of DLR on cine CMR’s image quality and precision in evaluating quantitative volume and strain was equal or superior to that of CS. DLR may replace CS for cine CMR.

Purpose: Diffusion models (DMs) excel in pixel-level and spatial tasks and are proven feature extractors for 2D image discriminative tasks when pretrained. However, their capabilities in 3D MRI discriminative tasks remain largely untapped. This study seeks to assess the effectiveness of DMs in this underexplored area.

Methods: We use 59830 T1-weighted MR images (T1WIs) from the extensive, yet unlabeled, UK Biobank dataset. Additionally, we apply 369 T1WIs from the BraTS2020 dataset specifically for brain tumor classification, and 421 T1WIs from the ADNI1 dataset for the diagnosis of Alzheimer’s disease. Firstly, a high-performing denoising diffusion probabilistic model (DDPM) with a U-Net backbone is pretrained on the UK Biobank, then fine-tuned on the BraTS2020 and ADNI1 datasets. Afterward, we assess its feature representation capabilities for discriminative tasks using linear probes. Finally, we accordingly introduce a novel fusion module, named CATS, that enhances the U-Net representations, thereby improving performance on discriminative tasks.

Results: Our DDPM produces synthetic images of high quality that match the distribution of the raw datasets. Subsequent analysis reveals that DDPM features extracted from middle blocks and smaller timesteps are of high quality. Leveraging these features, the CATS module, with just 1.7M additional parameters, achieved average classification scores of 0.7704 and 0.9217 on the BraTS2020 and ADNI1 datasets, demonstrating competitive performance with that of the representations extracted from the transferred DDPM model, as well as the 33.23M parameters ResNet18 trained from scratch.

Conclusion: We have found that pretraining a DM on a large-scale dataset and then fine-tuning it on limited data from discriminative datasets is a viable approach for MRI data. With these well-performing DMs, we show that they excel not just in generation tasks but also as feature extractors when combined with our proposed CATS module.

Purpose: To assess the effects of industry-developed deep learning reconstruction with super resolution (DLR-SR) on single-shot turbo spin-echo (SshTSE) images with thickness of 2 mm with DLR (SshTSE^2mm) relative to those of images with a thickness of 5 mm with DLR (SSshTSE^5mm) in the patients with pancreatic cystic lesions.

Methods: Thirty consecutive patients who underwent abdominal MRI examinations because of pancreatic cystic lesions under observation between June 2024 and July 2024 were enrolled. We qualitatively and quantitatively evaluated the image qualities of SshTSE^2mm and SshTSE^5mm with and without DLR-SR.

Results: The SNRs of the pancreas, spleen, paraspinal muscle, peripancreatic fat, and pancreatic cystic lesions of SshTSE^2mm with and without DLR-SR did not decrease in compared to that of SshTSE^5mm with and without DLR-SR. There were no significant differences in contrast-to-noise ratios (CNRs) of the pancreas-to-cystic lesions and fat between 4 types of images. SshTSE^2mm with DLR-SR had the highest image quality related to pancreas edge sharpness, perceived coarseness pancreatic duct clarity, noise, artifacts, overall image quality, and diagnostic confidence of cystic lesions, followed by SshTSE^2mm without DLR-SR and SshTSE^5mm with and without DLR-SR (P  <  0.0001).

Conclusions: SshTSE^2mm with DLR-SR images had better quality than the other images and did not have decreased SNRs and CNRs. The thin-slice SshTSE with DLR-SR may be feasible and clinically useful for the evaluation of patients with pancreatic cystic lesions.

Purpose: The purposes of this study were 1) to improve vessel visibility of our MR sequence by modifying k-space filling and 2) to verify the usefulness of applying artificial intelligence (AI) for volume isotropic simultaneous interleaved bright- and black-blood examination (VISIBLE) with compressed sensitivity encoding (CS) in autodetecting brain metastases.

Methods: We modified 3 sequences of VISIBLE (Centric, Reversed Centric, and Startup Echo 30). The Centric sequence is a prototype. The Reversed Centric filled the k-space in a reversed centric manner to improve vessel visibility. The Startup Echo 30 implemented dummy echoes to further improve vessel visibility. Vessel visibility was evaluated in one slice at the level of the centrum semiovale. The sensitivity, specificity, the area under the curve (AUC), and false positives of detecting brain metastases using AI were evaluated among 3 sequences. Statistical comparisons were performed using a one-way analysis of variance, followed by Friedman and Dunn’s multiple comparison tests.

Results: The number of visualized vessels was significantly lower in the Centric (39.3 ± 9.7, P < 0.05) and Reversed Centric (44.2 ± 9.8, P < 0.05) methods than in the magnetization-prepared rapid gradient echo (49.3 ± 9.1) but comparable in the Startup Echo 30 method (44.9 ± 8.8, P > 0.05). No significant differences existed in sensitivity, specificity, and AUC among the 3 methods. False positives achieved using the Reversed Centric method were significantly fewer (54 false positives) than those achieved using the Centric (85 false positives) and Startup Echo 30 (68 false positives) methods (P = 0.0092).

Conclusion: Vessel visibility was improved by modifying the k-space filling, which may reduce false positives. The AI model for VISIBLE with CS achieved good performance in autodetection of brain metastases. The AI model for VISIBLE with CS can help radiologists diagnose brain metastases in clinical practice.

Purpose: To reduce motion artifacts (MA) and noise in brain MRI using deep learning to promote clinical utility.

Methods: T1-weighted (T1W), T2-weighted (T2W), and fluid attenuated inversion recovery (FLAIR) images of the brain (including sagittal, coronal, and axial sections) of 20 healthy volunteers were collected using a 3.0T MR system. Simulated images with horizontal and vertical phase directions exhibiting varying white noise and MA (n = 115200) were created for each sequence and trained in deep learning (36000 pairs), validation (200 pairs), and testing (200 pairs, 2000 pairs) datasets. Images with MA and noise and images without MA and noise were included. A training model was constructed to remove noise and MA. The model’s ability to remove noise and MA was evaluated by the structural similarity index (SSIM) and peek signal to noise ratio (PSNR). The SSIM and PSNR between the ground-truth and output images were calculated (SSIMout, PSNRout), and the SSIM and PSNR between the ground-truth and input images were calculated (SSIMinp, PSNRinp). The ratio of SSIMinp to SSIMout was then evaluated as the improvement ratio of SSIM (IMPRs) and the ratio of PSNRinp to PSNRout as the improvement ratio of PSNR (IMPRp). In addition, 10 radio technologists performed visual evaluation.

Results: The SSIMout were >0.95 and 33 dB, respectively, for T1W, T2W, and FLAIR images with different contrasts. The mean value of SSIMinp was 0.72. Noise and MA removal effects were observed in images, with an average value of 72 dB. Visual evaluation revealed that the reduction effect in the output image was higher than that observed in the input image.

Conclusion: The method proposed herein, which uses separate training models for the T1W, T2W, and FLAIR sequences, is a valuable approach for removing MA and noise, independent of the imaging direction and artifact orientation. An improvement in image quality was achieved by generating a large amount of training data through simulation.

Purpose: To develop a new method to generate synthetic MR spectroscopic imaging (MRSI) data for training machine learning models.

Methods: This study targeted routine MRI examination protocols with single voxel spectroscopy (SVS). A novel model derived from pix2pix generative adversarial networks was proposed to generate synthetic MRSI data using MRI and SVS data as inputs. T1- and T2-weighted, SVS, and reference MRSI data were acquired from healthy brains with clinically available sequences. The proposed model was trained to generate synthetic MRSI data. Quantitative evaluation involved the calculation of the mean squared error (MSE) against the reference and metabolite ratio value. The effect of the location of and the number of the SVS data on the quality of the synthetic MRSI data was investigated using the MSE.

Results: The synthetic MRSI data generated from the proposed model were visually closer to the reference. The 95% confidence interval (CI) of the metabolite ratio value of synthetic MRSI data overlapped with the reference for seven of eight metabolite ratios. The MSEs tended to be lower in the same location than in different locations. The MSEs among groups of numbers of SVS data were not significantly different.

Conclusion: A new method was developed to generate MRSI data by integrating MRI and SVS data. Our method can potentially increase the volume of MRSI data training for other machine learning models by adding SVS acquisition to routine MRI examinations.

Purpose: To investigate the visibility of the lenticulostriate arteries (LSAs) in time-of-flight (TOF)-MR angiography (MRA) using compressed sensing (CS)-based deep learning (DL) image reconstruction by comparing its image quality with that obtained by the conventional CS algorithm.

Methods: Five healthy volunteers were included. High-resolution TOF-MRA images with the reduction (R)-factor of 1 were acquired as full-sampling data. Images with R-factors of 2, 4, and 6 were then reconstructed using CS-DL and conventional CS (the combination of CS and sensitivity conceding; CS-SENSE) reconstruction, respectively. In the quantitative assessment, the number of visible LSAs (identified by two radiologists), length of each depicted LSA (evaluated by one radiological technologist), and normalized mean squared error (NMSE) value were assessed. In the qualitative assessment, the overall image quality and the visibility of the peripheral LSA were visually evaluated by two radiologists.

Results: In the quantitative assessment of the DL-CS images, the number of visible LSAs was significantly higher than those obtained with CS-SENSE in the R-factors of 4 and 6 (Reader 1) and in the R-factor of 6 (Reader 2). The length of the depicted LSAs in the DL-CS images was significantly longer in the R-factor 6 compared to the CS-SENSE result. The NMSE value in CS-DL was significantly lower than in CS-SENSE for R-factors of 4 and 6. In the qualitative assessment of DL-CS images, the overall image quality was significantly higher than that obtained with CS-SENSE in the R-factors 4 and 6 (Reader 1) and in the R-factor 4 (Reader 2). The visibility of the peripheral LSA was significantly higher than that shown by CS-SENSE in all R-factors (Reader 1) and in the R-factors 2 and 4 (Reader 2).

Conclusion: CS-DL reconstruction demonstrated preserved image quality for the depiction of LSAs compared to the conventional CS-SENSE when the R-factor is elevated.

Purpose: To compare the utility of thin-slice fat-suppressed single-shot T2-weighted imaging (T2WI) with deep learning image reconstruction (DLIR) and conventional fast spin-echo T2WI with DLIR for evaluating pancreatic protocol.

Methods: This retrospective study included 42 patients (mean age, 70.2 years) with pancreatic cancer who underwent gadoxetic acid-enhanced MRI. Three fat-suppressed T2WI, including conventional fast-spin echo with 6 mm thickness (FSE 6 mm), single-shot fast-spin echo with 6 mm and 3 mm thickness (SSFSE 6 mm and SSFSE 3 mm), were acquired for each patient. For quantitative analysis, the SNRs of the upper abdominal organs were calculated between images with and without DLIR. The pancreas-to-lesion contrast on DLIR images was also calculated. For qualitative analysis, two abdominal radiologists independently scored the image quality on a 5-point scale in the FSE 6 mm, SSFSE 6 mm, and SSFSE 3 mm with DLIR.

Results: The SNRs significantly improved among the three T2-weighted images with DLIR compared to those without DLIR in all patients (P < 0.001). The pancreas-to-lesion contrast of SSFSE 3 mm was higher than those of the FSE 6 mm (P < 0.001) and tended to be higher than SSFSE 6 mm (P = 0.07). SSFSE 3 mm had the highest image qualities regarding pancreas edge sharpness, pancreatic duct clarity, and overall image quality, followed by SSFSE 6 mm and FSE 6 mm (P < 0.0001).

Conclusion: SSFSE 3 mm with DLIR demonstrated significant improvements in SNRs of the pancreas, pancreas-to-lesion contrast, and image quality more efficiently than did SSFSE 6 mm and FSE 6 mm. Thin-slice fat-suppressed single-shot T2WI with DLIR can be easily implemented for pancreatic MR protocol.

This article presents an overview of deep learning (DL) and its applications to function approximation for MR in medicine. The aim of this article is to help readers develop various applications of DL. DL has made a large impact on the literature of many medical sciences, including MR. However, its technical details are not easily understandable for non-experts of machine learning (ML).

The first part of this article presents an overview of DL and its related technologies, such as artificial intelligence (AI) and ML. AI is explained as a function that can receive many inputs and produce many outputs. ML is a process of fitting the function to training data. DL is a kind of ML, which uses a composite of many functions to approximate the function of interest. This composite function is called a deep neural network (DNN), and the functions composited into a DNN are called layers. This first part also covers the underlying technologies required for DL, such as loss functions, optimization, initialization, linear layers, non-linearities, normalization, recurrent neural networks, regularization, data augmentation, residual connections, autoencoders, generative adversarial networks, model and data sizes, and complex-valued neural networks.

The second part of this article presents an overview of the applications of DL in MR and explains how functions represented as DNNs are applied to various applications, such as RF pulse, pulse sequence, reconstruction, motion correction, spectroscopy, parameter mapping, image synthesis, and segmentation.

Purpose: Myelination-related MR signal changes in white matter are helpful for assessing normal development in infants and children. A rule-based myelination evaluation workflow regarding signal changes on T1-weighted images (T1WIs) and T2-weighted images (T2WIs) has been widely used in radiology. This study aimed to simulate a rule-based workflow using a stacked deep learning model and evaluate age estimation accuracy.

Methods: The age estimation system involved two stacked neural networks: a target network-to extract five myelination-related images from the whole brain, and an age estimation network from extracted T1- and T2WIs separately. A dataset was constructed from 119 children aged below 2 years with two MRI systems. A four-fold cross-validation method was adopted. The correlation coefficient (CC), mean absolute error (MAE), and root mean squared error (RMSE) of the corrected chronological age of full-term birth, as well as the mean difference and the upper and lower limits of 95% agreement, were measured. Generalization performance was assessed using datasets acquired from different MR images. Age estimation was performed in Sturge–Weber syndrome (SWS) cases.

Results: There was a strong correlation between estimated age and corrected chronological age (MAE: 0.98 months; RMSE: 1.27 months; and CC: 0.99). The mean difference and standard deviation (SD) were −0.15 and 1.26, respectively, and the upper and lower limits of 95% agreement were 2.33 and −2.63 months. Regarding generalization performance, the performance values on the external dataset were MAE of 1.85 months, RMSE of 2.59 months, and CC of 0.93. Among 13 SWS cases, 7 exceeded the limits of 95% agreement, and a proportional bias of age estimation based on myelination acceleration was exhibited below 12 months of age (P = 0.03).

Conclusion: Stacked deep learning models automated the rule-based workflow in radiology and achieved highly accurate age estimation in infants and children up to 2 years of age.