Xucheng Zhu

Background: Medical imaging datasets often suffer from class imbalance and limited availability of pathology-rich cases, which limit the performance of machine learning models for segmentation, classification, and vision–language tasks. We propose POWDR, a pathology-preserving outpainting framework for 3D MRI that uses real pathological regions as conditioning evidence while generating anatomically plausible surrounding tissue. Methods: Our approach leverages wavelet-domain conditioning to enhance high-frequency detail and mitigate blurring common in latent diffusion models. We introduce a random connected mask training strategy to reduce conditioning-induced collapse and improve diversity outside the lesion. POWDR is evaluated on brain MRI using BraTS datasets and extended to knee MRI to assess applicability beyond brain imaging. Results: Quantitative metrics (FID, MS-SSIM, LPIPS) were used to assess image realism. Random connected mask training improved diversity, reducing cosine similarity from 0.9947 to 0.9580 and increasing KL divergence from 0.00026 to 0.01494. To validate pathology preservation, we compared lesion overlap, volume, intensity, and morphology. For downstream segmentation, nnU-Net performance improved from 0.6992 to 0.7137 Dice after augmentation with 50 synthetic cases, representing a modest but statistically significant improvement (paired t-test, p = 0.016). Tissue volume analysis showed no significant differences for CSF and GM compared to real images, while WM volume was lower in synthetic images. Conclusions: POWDR provides a framework for generating diverse, pathology-preserving synthetic MRI data. The results suggest potential utility for data augmentation while maintaining clinically relevant lesion characteristics.

Medical imaging datasets often suffer from class imbalance and limited availability of pathology-rich cases, which constrains the performance of machine learning models for segmentation, classification, and vision-language tasks. To address this challenge, we propose POWDR, a pathology-preserving outpainting framework for 3D MRI based on a conditioned wavelet diffusion model. Unlike conventional augmentation or unconditional synthesis, POWDR retains real pathological regions while generating anatomically plausible surrounding tissue, enabling diversity without fabricating lesions. Our approach leverages wavelet-domain conditioning to enhance high-frequency detail and mitigate blurring common in latent diffusion models. We introduce a random connected mask training strategy to overcome conditioning-induced collapse and improve diversity outside the lesion. POWDR is evaluated on brain MRI using BraTS datasets and extended to knee MRI to demonstrate tissue-agnostic applicability. Quantitative metrics (FID, SSIM, LPIPS) confirm image realism, while diversity analysis shows significant improvement with random-mask training (cosine similarity reduced from 0.9947 to 0.9580; KL divergence increased from 0.00026 to 0.01494). Clinically relevant assessments reveal gains in tumor segmentation performance using nnU-Net, with Dice scores improving from 0.6992 to 0.7137 when adding 50 synthetic cases. Tissue volume analysis indicates no significant differences for CSF and GM compared to real images. These findings highlight POWDR as a practical solution for addressing data scarcity and class imbalance in medical imaging. The method is extensible to multiple anatomies and offers a controllable framework for generating diverse, pathology-preserving synthetic data to support robust model development.

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.

Purpose Deep learning-based reconstruction techniques, such as the prototype DL-Speed, have been developed to accelerate 3D T1-weighted imaging, but their clinical utility and impact on quantitative analysis have not been systematically validated. The purpose of this study is to evaluate the utility of a deep learning-based reconstruction method for substantially accelerating image acquisition without compromising quantitative image quality and morphometry. Methods Six healthy volunteers were scanned with 3D MPRAGE using acceleration factors ranging from 2 to 16. Image quality was assessed using the CAT12 Image Quality Rating (IQR), and morphometry with CAT12 global cortical thickness and total gray matter (GM) volume. In a clinical cohort of 40 patients, DL-Speed with 11-fold acceleration (1 min 10 s) was compared to conventional imaging (4 min 59 s). Head motion was quantified via total vector change (TVC) using volumetric navigators. Results In healthy volunteers, image quality remained within CAT12 Rank A at all acceleration levels; IQR followed a quadratic trend (R² = 0.98). Global cortical thickness decreased approximately linearly with acceleration (e.g., DLS2→DLS11: −0.068 mm [− 2.8%]; DLS2→DLS16: −0.140 mm [− 5.7%]). Total GM volume showed a shallow, monotonic decline (group means: 748.8 mL at DLS2 → 729.5 mL at DLS11 [− 2.6%] → 726.3 mL at DLS16 [− 3.0%]). In clinical cohort, DL-Speed significantly reduced head motion (TVC: 52.3 ± 9.4 mm vs. 140.4 ± 32.8 mm, p < 0.001) while maintaining acceptable image quality (IQR: 93.9 ± 1.0% vs. 94.8 ± 1.2%, p < 0.001). Bland–Altman analysis showed narrow limits of agreement (–2.49% to + 0.75%). Regarding cortical thickness and GM volume, we found significant correlations between conventional images and DL-Speed (thickness r = 0.97; GM r = 0.99; both p < 0.001). Conclusion DL-Speed enables about-one-minute 3D T1-weighted imaging that markedly reduces motion while preserving quantitative image integrity suitable for morphometry. Because image quality (IQR) follows a non-linear (quadratic) dependency whereas cortical thickness and GM volume decline nearly linearly, the data support a practical operating range (e.g., 6–11-fold acceleration) rather than a single optimal factor. Trial registration Not applicable.

Two-dimensional (2D) cine imaging is essential in routine clinical cardiac MR (CMR) exams for assessing cardiac structure and function. Traditional cine imaging requires patients to hold their breath for extended periods and maintain consistent heartbeats for optimal image quality, which can be challenging for those with impaired breath-holding capacity or irregular heart rhythms. This study aims to systematically assess the performance of a deep learning-based reconstruction (Sonic DL Cine, GE HealthCare, Waukesha, WI, USA) for accelerated cardiac cine acquisition. Multiple retrospective experiments were designed and conducted to comprehensively evaluate the technique using data from an MR-dedicated extended cardiac torso anatomical phantom (digital phantom) and healthy volunteers on different cardiac planes. Image quality, spatiotemporal sharpness, and biventricular cardiac function were qualitatively and quantitatively compared between Sonic DL Cine-reconstructed images with various accelerations (4-fold to 12-fold) and fully sampled reference images. Both digital phantom and in vivo experiments demonstrate that Sonic DL Cine can accelerate cine acquisitions by up to 12-fold while preserving comparable SNR, contrast, and spatiotemporal sharpness to fully sampled reference images. Measurements of cardiac function metrics indicate that function measurements from Sonic DL Cine-reconstructed images align well with those from fully sampled reference images. In conclusion, this study demonstrates that Sonic DL Cine is able to reconstruct highly under-sampled (up to 12-fold acceleration) cine datasets while preserving SNR, contrast, spatiotemporal sharpness, and quantification accuracy for cardiac function measurements. It also provides a feasible approach for thoroughly evaluating the deep learning-based method.

Deep learning–based cardiac cine sequences performed over three cardiac cycles led to substantially reduced acquisition time, while no differences in image quality and volumetric results compared with a standard clinical sequence could be detected.

Background: Ventricular volumetry using a short-axis stack of two-dimensional (D) cine balanced steady-state free precession (bSSFP) sequences is crucial in any cardiac magnetic resonance imaging (MRI) examination. This task becomes particularly challenging in children due to multiple breath-holds. Objective: To assess the diagnostic performance of accelerated 3-RR cine MRI sequences using deep learning reconstruction compared with standard 2-D cine bSSFP sequences. Material and methods: Twenty-nine consecutive patients (mean age 11 ± 5, median 12, range 1–17 years) undergoing cardiac MRI were scanned with a conventional segmented 2-D cine and a deep learning accelerated cine (three heartbeats) acquisition on a 1.5-tesla scanner. Short-axis volumetrics were performed (semi-)automatically in both datasets retrospectively by two experienced readers who visually assessed image quality employing a 4-point grading scale. Scan times and image quality were compared using the Wilcoxon rank-sum test. Volumetrics were assessed with linear regression and Bland–Altman analyses, and measurement agreement with intraclass correlation coefficient (ICC). Results: Mean acquisition time was significantly reduced with the 3-RR deep learning cine compared to the standard cine sequence (45.5 ± 13.8 s vs. 218.3 ± 44.8 s; P < 0.001). No significant differences in biventricular volumetrics were found. Left ventricular (LV) mass was increased in the deep learning cine compared with the standard cine sequence (71.4 ± 33.1 g vs. 69.9 ± 32.5 g; P < 0.05). All volumetric measurements had an excellent agreement with ICC > 0.9 except for ejection fraction (EF) (LVEF 0.81, RVEF 0.73). The image quality of deep learning cine images was decreased for end-diastolic and end-systolic contours, papillary muscles, and valve depiction (2.9 ± 0.5 vs. 3.5 ± 0.4; P < 0.05). Conclusion: Deep learning cine volumetrics did not differ significantly from standard cine results except for LV mass, which was slightly overestimated with deep learning cine. Deep learning cine sequences result in a significant reduction in scan time with only slightly lower image quality. Graphical

Purpose: To evaluate methods for quantification of pulmonary ventilation with ultrashort echo time (UTE) MRI. Methods: We performed a reproducibility study, acquiring two free-breathing ¹ H UTE lung MRIs on the same day for six healthy volunteers. The 1) 3D + t cyclic b-spline and 2) symmetric image normalization (SyN) methods for image registration were applied after respiratory phase-resolved image reconstruction. Ventilation maps were calculated using 1) Jacobian determinant of the deformation fields minus one, termed regional ventilation, and 2) intensity percentage difference between the registered and fixed image, termed specific ventilation. We compared the reproducibility of all four method combinations via statistical analysis. Results: Split violin plots and Bland-Altman plots are shown for whole lungs and lung sections. The cyclic b-spline registration and Jacobian determinant regional ventilation quantification provide total ventilation volumes that match the segmentation tidal volume, smooth and uniform ventilation maps. The cyclic b-spline registration and specific ventilation combination yields the smallest standard deviation in the Bland-Altman plot. Conclusion: Cyclic registration performs better than SyN for respiratory phase-resolved ¹ H UTE MRI ventilation quantification. Regional ventilation correlates better with segmentation lung volume, while specific ventilation is more reproducible.

Purpose: Three‐dimensional UTE MRI has shown the ability to provide simultaneous structural and functional lung imaging, but it is limited by respiratory motion and relatively low lung parenchyma SNR. The purpose of this paper is to improve this imaging by using a respiratory phase‐resolved reconstruction approach, named motion‐compensated low‐rank reconstruction (MoCoLoR), which directly incorporates motion compensation into a low‐rank constrained reconstruction model for highly efficient use of the acquired data.Theory and Methods: The MoCoLoR reconstruction is formulated as an optimization problem that includes a low‐rank constraint using estimated motion fields to reduce the rank, optimizing over both the motion fields and reconstructed images. The proposed reconstruction along with XD and motion state–weighted motion‐compensation (MostMoCo) methods were applied to 18 lung MRI scans of pediatric and young adult patients. The data sets were acquired under free‐breathing and without sedation with 3D radial UTE sequences in approximately 5 min. After reconstruction, they went through ventilation analyses. Performance across reconstruction regularization and motion‐state parameters were also investigated.Results: The in vivo experiments results showed that MoCoLoR made efficient use of the data, provided higher apparent SNR compared with state‐of‐the‐art XD reconstruction and MostMoCo reconstructions, and yielded high‐quality respiratory phase‐resolved images for ventilation mapping. The method was effective across the range of patients scanned.Conclusion: The motion‐compensated low‐rank regularized reconstruction approach makes efficient use of acquired data and can improve simultaneous structural and functional lung imaging with 3D‐UTE MRI. It enables the scanning of pediatric patients under free‐breathing and without sedation.

Manual prescription of the field of view (FOV) by MRI technologists is variable and prolongs the scanning process. Often, the FOV is too large or crops critical anatomy. We propose a deep learning framework, trained by radiologists’ supervision, for automating FOV prescription. An intra-stack shared feature extraction network and an attention network are used to process a stack of 2D image inputs to generate scalars defining the location of a rectangular region of interest (ROI). The attention mechanism is used to make the model focus on a small number of informative slices in a stack. Then, the smallest FOV that makes the neural network predicted ROI free of aliasing is calculated by an algebraic operation derived from MR sampling theory. The framework’s performance is examined quantitatively with intersection over union (IoU) and pixel error on position and qualitatively with a reader study. The proposed model achieves an average IoU of 0.867 and an average ROI position error of 9.06 out of 512 pixels on 80 test cases, significantly better than two baseline models and not significantly different from a radiologist. Finally, the FOV given by the proposed framework achieves an acceptance rate of 92% from an experienced radiologist.