Magnetic Resonance in Medicine

Sodium (23Na) magnetic resonance imaging (MRI) provides valuable metabolic information, but it is limited by a low signal-to-noise ratio (SNR) and long acquisition times. To overcome these challenges, we present a Deep Image Prior (DIP)-based framework that combines anatomically guided proton (1H) MRI and metabolically guided 23Na MRI denoising via a fused proton-sodium prior within a directional total variation (dTV) regularization scheme. The DIP-Fusion approach minimizes a variational loss function combining data fidelity, fused dTV regularization, gradient consistency, and bias-field correction to reconstruct sodium images. MRI data were acquired from healthy volunteers and breast cancer patients. Healthy datasets were retrospectively undersampled at multiple factors, and fully sampled scans served as the ground truth. Patient datasets acquired for clinical purposes were reconstructed using the baseline DIP and the proposed DIP-Fusion methods. Sodium images were reconstructed using sum-of-squares (SoS) and adaptive combined (ADC) coil combination methods. We evaluated reconstruction performance using quantitative image quality metrics, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), mean squared error (MSE), learned perceptual image patch similarity (LPIPS), feature similarity index (FSIM), and Laplacian focus. In healthy volunteers, DIP-Fusion outperformed state-of-the-art reconstruction techniques across all undersampling factors. In patient datasets, DIP-Fusion demonstrated superior performance compared with baseline DIP, achieving improved structural fidelity and sodium-specific signal preservation. These results demonstrate the potential for robust, highquality sodium MRI reconstruction under accelerated acquisition, which could lead to reduced scan times and enhanced clinical feasibility.

Chemical shift-encoded magnetic resonance imaging using high-resolved 3D Dixon techniques enables the non-invasive and radiation-free assessment of whole-body adipose tissue and ectopic fat distribution. Automatic deep learning-based segmentation of metabolically relevant adipose tissue compartments and ectopic fat deposits in parenchymal tissue is the most important image processing step for the quantification of adipose tissue volumes and ectopic fat percentages from whole-body imaging. This work presents a segmentation model dedicated to the segmentation of 19 metabolically relevant adipose tissue compartments and ectopic fat deposits from whole-body Dixon MRI. The trained segmentation model is available upon request. Related post-processing routines to compute volumes and fat percentages are publicly available: https://github.com/tobihaui/WholeBodyATQuantification.

PurposeAqueous humor drains fluid from the eye not only via the conventional pathway through the trabecular meshwork and Schlemms canal, but also within the eye is known to occur via pathways through the posterior chamber and optic nerve to the cerebrospinal fluid (CSF) surrounding the optic nerve. The mechanism is poorly understood, and non-invasive method for evaluation in living humans has not been established. We previously showed that eye drops containing O-17-labeled water (H217O) distribute in the anterior chamber but not the vitreous. This study aimed to evaluate the distribution of H217O in the CSF along the optic nerve. MethodsFive ophthalmologically normal participants (20-31 years, all females) were selected from a previous prospective study based on 1H MR images of the eyes that included the optic nerve. They received eye drops of 10 mol% H217O in their right eye. Dynamic image time series was created by normalizing the signal of each 1H-T2WI by the pre-drop average signal. Region-of-interest analyses were performed for signal changes in the anterior chamber, vitreous, and CSF. ResultsIn the quantitative evaluation, the normalized intensity in the anterior chamber and CSF was significantly lower than that in the pre-drop signal (anterior chamber: 0.78 {+/-} 0.07, p < 0.005; CSF: 0.89 {+/-} 0.07, p < 0.05). No distribution was identified in the vitreous. Qualitatively, the distribution of H217O in the anterior chamber was detected in all five participants and in the CSF of four participants (80%). ConclusionH217O eye drops were distributed in the anterior chamber and CSF, but not in the vitreous. These findings suggest that the visualization of aqueous humor outflow, not via the Schlemms canal, may contribute to ocular fluid homeostasis, including the ocular glymphatic system.

BackgroundDynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) enables non-invasive characterization of carotid atherosclerotic plaque. PurposeTo evaluate the performance and reproducibility of a simplified DCE-MRI quantification method for carotid plaque assessment. MethodsT1-weighted black-blood DCE-MRI of the carotid arteries at 3T was performed at baseline and after six months in patients with mild-to-moderate atherosclerotic lesions in a pilot placebo-controlled randomized trial evaluating the effects of low-dose (0.5mg daily) colchicine therapy on carotid plaque volume. DCE-MRI signal intensity was measured in manually drawn regions of interest in the plaque core, remote non-atherosclerotic vessel wall, and skeletal muscle. Peak signal intensities were normalized to skeletal muscle signal in the same slice. ResultsIn patients (n=28, median [interquartile range] age 72 [64-74] years, 36% female, n=13/15 colchicine/placebo), normalized peak signal intensity was higher in the plaque core than in the remote vessel wall at both baseline (3.5 [2.3-4.1] vs 2.1 [1.7-2.5], p<0.001) and follow-up (3.2 [2.5-4.4] vs 2.0 [1.7-2.5], p<0.001). Measurements did not differ between baseline and follow-up for all patients (0.7{+/-}0.7 for plaque core, 0.6{+/-}0.4 for remote vessel wall, p>0.80 for both) nor between colchicine intervention and placebo control (p>0.35 for either region). ConclusionsNormalised peak signal intensity on DCE-MRI was consistently higher in the carotid plaque core than in the remote vessel wall, showed excellent reproducibility in both regions over six months, and was not altered by colchicine treatment. This simplified, muscle-normalised approach may facilitate future studies exploring DCE-MRI measures potentially related to plaque vulnerability.

Magnetic Resonance Spectroscopy Imaging (MRSI) offers spatially-resolved, neurometabolic information, acquired non-invasively at whole-brain scales from human subjects. Analysis of MRSI however, is extremely challenging. The metabolic information is highly convolved, and sparsely distributed across millions of spatial-spectral datapoints, allowing for little direct human interpretation. Conversely, the overall low signal-to-noise with high-intensity artifacts can confound unsupervised machine learning approaches. These technical barriers have left much of the potential of MRSI unrealized. We acquired MRSI data from 4 human subjects with a diagnosis of multiple sclerosis (MS), incorporating experimental design into an informed machine learning approach. MRSI acquisitions were registered to anatomical MRI to label 105k spectra from brain tissue and 162 spectra from white matter hyperintensities (WMHs), an imaging biomarker associated with MS lesions. Spectral labels were then used in contrastive principal component analysis (cPCA) to filter artifacts and background features in the MRSI data from lesion salient features and clustered into statistically significant states based on features that could be interpreted from the original data. Our approach renders MRSI data into testable representations of neurometabolism, enabling the method for fundamental and clinical research. Graphical AbstractAnalysis workflow for neurometabolic profiling of MS lesions. MRSI and anatomical MRI is acquired and processed in parallel for spectral data and anatomical labels. Spectra are then labeled and separated into experimental vs background data for contrastive PCA. Spectra are clustered for similarity, further labeled, and projected onto a brain atlas for a neurometabolic view. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=71 SRC="FIGDIR/small/26346248v1_ufig1.gif" ALT="Figure 1"> View larger version (28K): org.highwire.dtl.DTLVardef@21a1eorg.highwire.dtl.DTLVardef@e312org.highwire.dtl.DTLVardef@3bce70org.highwire.dtl.DTLVardef@6e56ae_HPS_FORMAT_FIGEXP M_FIG C_FIG

Background and Purpose: Magnetic resonance imaging (MRI) for radiation therapy treatment planning is currently being used in many anatomical sites to better visualize soft tissue landmarks, a technique known as an MRI simulation. A core component of modern MRI simulation configurations are the use of external laser positioning systems (ELPS) to help set up the patient. Though necessary for accurate and reproducible patient setup, the ELPS, if left on during imaging, may interfere negatively with image quality due to leaking electronic noise, of which MRI is sensitive to. It is currently unknown whether this leakage of electronic noise may further affect quantitative values derived from clinically employed relaxometric, diffusion, and fat fraction sequences. Therefore, in this study, we aim to characterize the impact of MRI simulation lasers on general image quality and quantitative imaging accuracy. Materials and Methods: First, a cine acquisition was used to visualize the real-time changes in image signal-to-noise ratio (SNR) from when the ELPS was deactivated to activated. To validate this effect quantitatively, the SNR was measured using the American College of Radiology (ACR) recommended protocol in a homogeneous phantom with the integrated body, 18-channel UltraFlex small, 18-channel UltraFlex large, 32-channel spine, and 16-channel shoulder coils. Next, a geometric distortion algorithm was tested in two vendor-provided phantoms while using the integrated body coil and the ACR Large Phantom protocol was tested. Finally, a series of quantitative MRI scans were performed using a CaliberMRI Model 137 Mini Hybrid phantom to validate quantitative T1, T2, and ADC while a Calimetrix PDFF-R2* phantom was used for quantitative PDFF and R2*. All scans were performed with both the ELPS both deactivated and activated. Results: Visible electronic noise artifacts were seen when using the integrated body coil when the ELPS was activated on the cine acquisition which led to a four-fold decrease in SNR using the ACR protocol. This SNR drop was not seen when using the remaining tested coils. The automatic fiducial detection algorithm was affected negatively by ELPS activation leading to misidentification when identified perfectly with the ELPS deactivated. Degradation in image intensity uniformity, percent signal ghosting, and low contrast object detectability was seen during ACR Large Phantom testing using the 20-channel Head/Neck coil. Concordance across quantitative MRI values was similar when the ELPS was both deactivated and activated while a consistent increase in standard deviation inside the ADC vials was seen when the ELPS was activated. Discussion: The extra noise induced from the activation of the ELPS during imaging should be avoided due to its potential to unnecessarily increase image noise. This is particularly true when conducting mandatory quality assurance testing for image quality and geometric distortion which utilize the integrated body coil which is most susceptible to ELPS-induced noise. Clear clinical guidelines should be implemented to make this issue known to the MRI technologists, physicists, and other relevant staff using an MRI with a supplementary ELPS for patient alignment.

BackgroundBreast cancer-related lymphedema (BCRL) is a common complication following breast cancer treatment. While lymphoscintigraphy is considered the diagnostic gold standard, it is unsuitable for routine periodic monitoring or assessment of treatment efficacy. Shear wave elastography (SWE) offers a possible alternative, but traditional modes of operation limit its potential. Proposed SolutionsThe Holder-Optimized Elastography (HOE) method is introduced to eliminate pressure issues introduced by manual operation of ultrasound probes by stabilizing them above the cutis. MethodsThe HOE method was used to acquire ARFI images of high-velocity areas (HVAs, with shear wave velocity greater than 7 m/s) in limbs with and without BCRL (as confirmed and characterized by lymphoscintigraphy) in two cohorts of 15 and 125 patients. ResultsThe HOE method enabled ARFI elastography to directly and consistently visualize the effects caused by both obstructed lymphatic vessels and intraluminal lymphatic fluid as HVAs, whereas traditional hand-held methods did not. Inter-limb differences in HVA burden showed moderate diagnostic performance for detecting BCRL and grading obstruction with modest sensitivity. However, there was systematic underestimation of both early and confluent advanced lesions. ConclusionHOE-based HVA imaging has potential for rapid and non-invasive monitoring of lymphedema course and treatment response and may serve as a useful adjunct to existing diagnostic tools for BCRL. However, further technical refinements and quantitative analytic methods will be required to fully exploit the richer SWV information provided by HOE and to enhance the diagnostic utility of HVAs. Summary StatementThe Holder-Optimized Elastography method ("HOE" method) increases the diagnostic capability of ARFI elastography for breast cancer-related lymphedema, allowing for the non-invasive detection of some lymphatic obstructions but not all. Key ResultsThe Holder-Optimized Elastography (HOE) method revealed the effects caused by fluid-filled lymphatic vessels as "High-Velocity Areas" (HVAs), which are difficult to detect by conventional methods. HVA counts for detecting lymphedema (any obstruction vs. no obstruction) showed high specificity (0.86-1.00) but low sensitivity (0.57-0.67). Conversely, HVA counts for staging lymphedema (i.e. total vs. partial obstruction) showed high sensitivity (up to 1.00) but low specificity (0.48-0.66). The inter-limb difference of HVAs counted in whole-limb scans between affected and unaffected limbs (aka, the "Global Mean Difference") provided the most balanced diagnostic performance (sensitivity 0.67-0.79, specificity 0.88-0.89).

Handheld ultrasound devices have revolutionized point-of-care diagnostics, but their effectiveness remains limited by operator dependency and the need for specialized training. This paper presents an intelligent guidance and diagnostic assistance system for the handheld wireless ultrasound device, enabling automated carotid artery and thyroid examinations through handheld operation. Drawing inspiration from the Actor-Critic framework, we implement a simulation-based reinforcement learning approach for real-time probe navigation toward standard anatomical views. The system integrates YOLOv8n-based detection networks for carotid plaque and thyroid nodule identification, achieving real-time inference at 30 frames per second. Furthermore, we propose a hybrid measurement approach combining UNet segmentation with the Snake algorithm for precise biometric quantification, including carotid intima-media thickness (IMT), lumen diameter, and lesion dimensions. Experimental validation on clinical datasets demonstrates that the proposed system achieves 91.2% accuracy in standard plane acquisition, 87.5% mean average precision (mAP) for plaque detection, and 89.3% mAP for nodule identification. Measurement results show excellent agreement with expert sonographers, with IMT measurements exhibiting a mean absolute difference of 0.08 mm. These findings demonstrate the feasibility of intelligent handheld ultrasound examination, significantly reducing operator dependency while maintaining diagnostic accuracy comparable to experienced clinicians.

Deep-learning based super-resolution has shown promise for enhancing the spatial resolution of brain magnetic resonance images, which may help visualize small anatomical structures more clearly. However, when only limited training data are available, it remains uncertain which model assessment method provides the most reliable estimate of out-of-sample performance. In this study, three widely used assessment strategies (three-way holdout, k-fold cross-validation, and nested cross-validation) were compared for evaluating the performance of such models in small datasets. Across 30 iterations, we randomly selected subsets of 20 T2-weighted images from the 1,113 scans of the Human Connectome Project. Each subset was used to train a model and estimate performance using the three methods. The ground truth error was computed from the remaining images. The assessment error is the difference between the estimated error and the ground truth error. The median assessment errors were 0.11,- 0.13, - 0.32 for three-way holdout, k-fold cross-validation, and nested cross-validation, respectively, with the cross-validation methods showing considerably smaller dispersions. Nested cross-validation selected fewer epochs, indicating more conservative model selection, but required substantially greater computational time, over three times longer than three-way holdout and more than twenty times longer than k-fold cross-validation. Our findings suggest that k-fold cross-validation offers the most favourable balance between accuracy, stability, and computational feasibility in small datasets. Further research is needed to determine how model complexity, dataset size, and the number of cross-validation folds influence assessment accuracy.

Rationale and ObjectivesContrast-enhanced (CE) MRI provides clear corticomedullary contrast for renal compartment delineation but may be contraindicated or undesirable in routine practice. We aimed to enable automated extraction of renal imaging biomarkers from routine non-contrast-enhanced (NCE) T1-weighted MRI by transferring CE-derived compartment labels. Materials and MethodsThis retrospective single-center study (January 2017 to December 2021) included 200 participants with paired arterial-phase CE and NCE T1-weighted MRI. Cortex, medulla, and sinus were manually segmented on CE MRI and rigidly transferred to NCE MRI to provide voxel-level reference labels. A hierarchical 3D Deep Neural Patchworks model was trained on 100 examinations (90 training/10 validation) and evaluated on an independent test set of 100 examinations using the transferred CE masks on NCE as reference. Performance was assessed using Dice similarity of segmentations and biomarker agreement using volumes and surface areas (Pearson/Spearman, MAE, Lins CCC, and Bland-Altman). ResultsWhole-kidney segmentation Dice was 0.950 (left) and 0.953 (right). Total kidney volume showed high agreement with minimal bias (MAE 8.76 mL, 2.5% of mean; CCC 0.983; bias -1.56 mL; 95% limits of agreement -28.81 to 25.69 mL). Cortex volume was modestly overestimated and medulla volume underestimated, shifting predicted compartment fractions toward cortex (74.7% vs. 72,1% in ground truth; medulla 21.5% vs. 24.3%; sinus 3.8% vs. 3.6%. Sinus volume maintained high concordance despite higher Dice dispersion. Surface area was systematically underestimated with low concordance. ConclusionCE-supervised knowledge transfer enables accurate, well-calibrated kidney volumetry from routine NCE MRI and supports contrast-free renal biomarker extraction. Surface area estimation remains challenging. Take-home MessagesO_LICE-supervised label transfer enables accurate, well-calibrated contrast-free kidney volumetry on routine non-contrast T1-weighted MRI. C_LIO_LICompartment volumetry is feasible but shows systematic cortex overestimation and medulla underestimation; surface area remains non-interchangeable due to boundary uncertainty. C_LI

BackgroundMotor threshold (MT) estimation is fundamental to transcranial magnetic stimulation (TMS), guiding individualized stimulation intensity in research and therapy. Conventional methods such as the 5-out-of-10 rule require many stimuli, while adaptive approaches like Parameter Estimation by Sequential Testing (PEST) improve efficiency but can exhibit poor convergence under certain conditions. ObjectiveThis study introduces the Bayesian Uncertainty Dynamic Algorithm for Parameter Estimation by Sequential Testing (BUDAPEST), a Bayesian adaptive method for fast, accurate MT estimation with user-controlled uncertainty. The aims were to validate its accuracy in simulations and human data, promote usability through a MATLAB-based graphical interface, and evaluate experimental utility through resting and active MT comparisons and session-to-session reliability. MethodsBUDAPEST infers MT from binary MEP responses using sequential Bayesian updating and terminates when a user-defined uncertainty threshold is reached. Performance was evaluated in 10,000 virtual simulations and in human rMT and aMT measurements across two sessions per subject, including 3x5 cortical motor mapping to assess physiological spatial patterns. ResultsIn simulations, BUDAPEST achieved a mean absolute error of 1.9% MSO within ~10 pulses using a 2% uncertainty criterion while avoiding PEST misestimations. In human data, MT estimates were accurate within {+/-}4% MSO and robust to initialization; rMT showed strong session-to-session reliability (r = 0.78), whereas aMT exhibited greater variability. Motor mapping revealed coherent excitability gradients centered on the hotspot. ConclusionBUDAPEST enables rapid, reliable, and uncertainty-controlled MT estimation while reducing procedure time and participant burden. The accompanying GUI facilitates immediate adoption in research and clinical TMS environments. HighlightsO_LIIntroduces BUDAPEST, a Bayesian uncertainty-aware algorithm for rapid and reliable TMS motor threshold estimation. C_LIO_LIAchieves accurate MT estimates ({approx}2% MSO error) in ~10 pulses with user-controlled trade-offs between precision and procedure duration. C_LIO_LIDemonstrates robust performance in simulations and human data, with strong resting MT reliability and an open-source GUI enabling immediate adoption. C_LI

Quantifying muscle health at scale has been limited by the difficulty of segmenting individual muscles on MRI. We developed an automated 3D deep-learning framework that segments 20 bilateral hip and thigh muscles from Dixon MRI, enabling muscle level quantification of volume and relative fat fraction (rFF). Applied to 10,840 baseline and 2,766 longitudinal UK Biobank scans, this framework supports population-scale phenotyping across demographic, metabolic and treatment exposures. Segmentation accuracy was robust, and increased with muscle size. Men had greater muscle volumes, whereas women showed consistently higher rFF. Fat infiltration was highest in postural and pelvic-stabilising muscles and lowest in the quadriceps, revealing pronounced anatomical heterogeneity. Over two years, most muscles showed small but consistent volume declines, with losses more uniform in men and more heterogeneous in women; rFF increased more prominently in women, suggesting early compositional deterioration. In T2D, men showed widespread volume loss and elevated rFF, whereas women showed minimal volume loss and heterogeneous fat changes, revealing sex-specific disease signatures. Automated muscle-specific MRI phenotyping resolves structural and compositional changes obscured by compartment-level measures and provides a scalable platform for population-level studies of musculoskeletal ageing, metabolic disease, and therapeutic response.

Abstract Background: Conventional MRI cannot reliably distinguish radiation necrosis (RN) from recurrent metastasis after cranial radiotherapy, as both can show similar enhancement despite different biology. We tested whether these entities are mechanically non-equivalent in vivo and separable by MRE-derived viscoelastic metrics and perilesional interface-instability features. Methods: In a prospective, histopathology-anchored cohort, 11 post-radiotherapy enhancing lesions were classified as RN (n=3) or recurrent/progressive tumor (n=8). MRE was acquired at 3.0 T with single-frequency 60-Hz excitation to derive storage modulus (G'), loss modulus (G''), and complex shear modulus magnitude (|G*|). Co-primary endpoints were median tumor G' and |G*|, each tested one-sided (RN > tumor) with Holm correction across the two co-primary tests. Median tumor G'' was tested two-sided. A prespecified secondary 6-endpoint family (absolute and tumor/NAWM-normalized G', G'', and |G*|) was analyzed with Benjamini-Hochberg FDR control. Exploratory instability mapping in a 0- 6 mm peritumoral shell generated interface-topology metrics, including convexity. Results: Absolute tumor-core medians were higher in RN than tumor for |G*| (1.79 vs 1.32 kPa; Cliff's {delta} = 0.67; q = 0.10), G' (1.62 vs 1.09 kPa; {delta} = 0.50; q = 0.14), and G'' (0.81 vs 0.46 kPa; {delta} = 0.75; q = 0.10). NAWM normalization improved separation: tumor/NAWM |G*| (2.26 vs 1.41; {delta} = 0.92; q = 0.04) and tumor/NAWM G'' (2.67 vs 0.87; {delta} = 1.00; q = 0.04) were FDR-significant. Convexity also differentiated RN from tumor (0.49 vs 0.36; {delta} = 1.00; MWU p = 0.01). Conclusions: Tumor/NAWM G'', tumor/NAWM |G*|, convexity, and tumor G'' emerged as the strongest candidate features, indicating that RN is mechanically harder and more dissipative than recurrent metastasis. Signal strength was high (Cliff's {delta} up to 1.00) but should be interpreted cautiously given sample size. Exploratory analyses further suggest that instability mapping captures biologically relevant interface behavior. These findings support a mechanics-based RN-versus-recurrence framework and justify prespecified, preregistered external validation.

Artificial intelligence (AI) based segmentation has many medical applications but limited curated datasets challenge model training; this study compares the impact of dataset annotation quality and quantity on whole liver AI segmentation performance. We obtained 3,089 abdominal computed tomography scans with whole-liver contours from MD Anderson Cancer Center (MDA) and a MICCAI challenge. A total of 249 scans were withheld for testing of which 30, MICCAI challenge data, were reserved for external validation. The remaining scans were divided into mixed-curation and highly-curated groups, randomly sampled into sub-datasets of various sizes, and used to train 3D nnU-Net segmentation models. Dice similarity coefficients (DSC), surface DSC with 2mm margins (SD 2mm), the 95th percentile of Hausdorff distance (HD95), and 2D axial slice DSC (Slice DSC) were used to evaluate model performance. The highly curated, 244-scan model (DSC=0.971, SD 2mm=0.958, HD95=2.98mm) performed insignificantly different on 3D evaluation metrics to the mixed-curation 2,840-scan model (DSC=0.971 [p>.999], SD 2mm=0.958 [p>.999], HD95=2.87mm [p>.999]). The 710-scan mixed-curation (Slice DSC=0.929) significantly outperformed the highly curated, 244-scan model (Slice DSC=0.923 [p=0.012]) on the 30 external scans. Highly curated datasets yielded equivalent performance to datasets that were a full order of magnitude larger. The benefits of larger, mixed-curation datasets are evidenced in model generalizability metrics and local improvements. In conclusion, tradeoffs between dataset quality and quantity for model training are nuanced and goal dependent.

We propose a lesion-centric phenotype learning pipeline for interpretable breast ultrasound (BUS). Predicted lesion masks are used for mask-weighted pooling of segmentation-encoder latents, producing compact embeddings that suppress background influence; a lightweight calibration step improves cross-dataset consistency. We cluster embeddings to discover latent phenotypes and relate phenotype structure to morphology descriptors (compactness, boundary sharpness). On BUSI and BUS-UCLM with external testing on BUS-BRA, lesion-centric pooling and calibration improve separability and enable strong malignancy probing (AUC 0.982), outperforming radiomics and a standard CNN baseline. A simple rule-gated generator further improves BI-RADS-style descriptor consistency on difficult cases.

Dense breast tissue diminishes the sensitivity of mammographic screening and is a key cancer risk factor, which motivates accurate segmentation under scarce and expensive expert annotations in the medical imaging domain. Here, we benchmark the effect of backbone architecture, self-supervised pre-training (SSL), fine-tuning strategy, and loss design for dense-tissue segmentation on a small expert-labeled dataset (596 images) and an in-domain unlabeled corpus (20, 000 images), reflecting the lack of large public pixel-level density datasets. CNNs (EfficientNet, Xception, nnUNet) clearly outperform transformer and Medical-SAM2 models, and full or layer-wise fine-tuning reliably exceeds parameter-efficient updates. Generic image-only SSL (MIM, SimCLR, Barlow Twins) often yields negligible or negative gains over ImageNet initialization, whereas a simple multi-view contrastive SSL and a hybrid segmentation-density loss provide the best accuracy and calibration (e.g., MAE from 14.8% to 11.8%, Spearman with the four BI-RADS breast density categories from 0.42 to 0.51 on VinDr). We also quantify GPU hours for different SSL and fine-tuning choices, showing that only a small set of protocols, such as EfficientNet with multi-view SSL, hybrid loss, and full fine-tuning, offers favorable accuracy-efficiency trade-offs. These findings provide practical defaults for annotation-limited mammography studies and support compute-conscious deployment of automatic breast density assessment in web-based screening workflows.

Radiogenomics enables the non-invasive characterisation of the genomic and molecular properties of tumours, with epidermal growth factor receptor (EGFR) mutations in non-small cell lung cancer (NSCLC) being one of the most investigated applications. In this study, we evaluate radiomics, contrastive learning, and convolutional deep learning approaches to predict the EGFR mutation status from chest Computed Tomography (CT) images using the TCIA Radiogenomics dataset (n=115). Our results, using 10-fold cross validation, demonstrate the capacity of imaging models to predict mutation status from CT data in a manner consistent with existing literature. Among the evaluated methods, models integrating radiomic with clinical features achieved the best performance, with an AUC of 0.790 and AUPRC of 0.517, outperforming both contrastive learning (AUC=0.787) and convolutional architectures (AUC=0.763). Beyond methodological comparisons, we discuss the challenges related to clinical translation. Specifically, we contrast radiogenomics with conventional tissue biopsies, and identify scenarios where radiogenomics might be useful, either independently or in conjunction with other existing diagnostic technologies. Together these findings evidence the potential utility of radiogenomics EGFR models and provide direct architecture comparisons on the same dataset.

Background and ObjectivesWhite matter hyperintensities (WMH) of presumed vascular origin are a neuroimaging hallmark of cerebral small vessel disease (CSVD). Their spatial heterogeneity may reflect different clinical phenotypes. Most prior studies relied on principal component analysis to characterise such heterogeneity, which has limited ability to stratify individuals into discrete and interpretable WMH subtypes. We therefore propose a data-driven framework to identify WMH spatial subtypes, characterise their demographic and clinical profiles, and investigate their predictive value for future WMH progression. MethodsWe analysed MRI scans from 63,338 individuals across 4 major cohorts (internal data): ADNI3, Insight46, SABRE and UK Biobank (UKB), and validated our findings in the OASIS-3 dataset (n=844). WMH were automatically segmented and regionally quantified using a 36-region bullseye framework. Clustering was applied to the relative regional distributions of WMH. A stability-based approach was used to identify robust WMH subtypes. Their associations with 19 risk factors of interest were analysed using multivariable regression. In a subset with follow-up MRI scans (internal: n=5,274, OASIS-3: n=182), we evaluated the predictive value of these subtypes combined with other volumetric or spatial WMH variables for WMH progression. ResultsFive WMH location patterns with different lesion burden and spatial distribution were identified (stability score 0.946) and reproduced in OASIS-3. These patterns showed distinct associations with demographic, vascular, metabolic, inflammatory and genetic risk factors. Higher-burden patterns were independently associated with older age, higher blood pressure, diabetes and smoking, indicating a gradient of vascular risk across spatial subtypes. WMH location patterns were largely preserved over 18-30 months, with most individuals remaining within the same pattern (71.5%). While global baseline WMH volume remained a strong predictor of future WMH progression (balanced accuracy 0.693, 95% CI: 0.664-0.723), models including baseline regional WMH volumes consistently outperformed other candidates (best balanced accuracy 0.737, 95% CI: 0.706-0.764). DiscussionWe presented a robust and scalable framework for spatial WMH phenotyping. We discussed clinical and prognostic implications of the spatial subtypes beyond total lesion burden. Our findings supported the value of WMH spatial characterisation in stratifying risk that may help guide personalised approaches to managing CSVD.

Digital breast tomosynthesis (DBT) is a powerful imaging modality that allows for improved lesion visibility, characterization, and localization compared to conventional two-dimensional digital mammography. DBT has been increasingly adopted in screening and diagnostic settings globally, particularly for women with dense breast tissue where tissue overlap presents a significant diagnostic challenge. Here we describe DBT-2026, a real world imaging dataset with 558 DBT exams from 558 patients with breast imaging reporting and data system (BI-RADS) scores of 0, 1, or 2. Each case contains one DBT examination in combination with expert annotations and free-text radiology reports that describe the radiological findings, produced in routine clinical practice. To protect patient privacy, all images and reports have been de-identified. The dataset is made freely available to researchers for non-commercial projects to facilitate and encourage research in breast cancer imaging.

Current standard of care imaging practices cannot reliably differentiate among certain renal tumors such as benign oncocytoma and clear cell renal cell carcinoma (RCC), and between low and high grade RCCs. Previous work has explored using deep learning, radiomics, and texture analysis to predict renal tumor subtypes and differentiate between low and high grade RCCs with mixed success. To further this work, large diverse datasets are needed to improve model performance and provide strong evaluation sets. In this work, a dataset of 831 multi-phase 3D CT exams was curated. Each exam contains up to three contrast-enhanced CT phases. Tumor outlines or bounding boxes were annotated and registered to the image volumes. The pathology results for each tumor and relevant patient metadata are also included.