European Radiology Experimental最新文献

Fetal MRI has emerged as a crucial supplement to prenatal ultrasonography in the evaluation of the developing brain and in identifying congenital defects and minor developmental malformations. While fetal brain MRI interpretation has always depended on visual examination of signal properties and morphology, images can provide quantitative information that could be missed or hidden from the human eye. Radiomics allows for characterizing tissue characteristics and heterogeneity by extracting quantitative information from imaging data. In this narrative review, after summarizing the technical foundations of fetal MRI radiomics (acquisition, preprocessing, segmentation, feature extraction and types, machine learning models, feature reproducibility and quality), we consider the following major clinical applications: brain development assessment and phenotyping; Chiari II malformation and brain edema phenotype; isolated ventriculomegaly and prediction of its persistence; and prognosis and neurodevelopmental outcome prediction. MRI radiomics presents a promising technique to improve the assessment of the fetal brain. Larger multicenter studies with standardized protocols are essential to improve generalizability and reduce variability. Combining radiomics with deep learning could enhance performance and interpretability, while biological validation, linking features to known tissue properties, will help confirm clinical relevance. RELEVANCE STATEMENT: Despite its early stage, MRI radiomics offers a new, data-driven lens to evaluate fetal brain development. By revealing subtle imaging patterns not visible to the eye, it may eventually support more accurate diagnosis, risk stratification, and personalized care. KEY POINTS: Fetal MRI adds value beyond ultrasound in the prenatal setting. Radiomics reveals hidden imaging features. Radiomics enhances diagnosis and prognosis in fetal brain assessment. Large multicenter studies are needed.

Objective: Chronic pulmonary embolism (CPE) and chronic thromboembolic pulmonary hypertension (CTEPH) are challenging to diagnose, with delayed detection increasing mortality. We evaluated the performance of a convolutional neural network (CNN) in identifying these conditions from computed tomography pulmonary angiography (CTPA)-derived maximum intensity projection (MIP) images using a novel approach including proximal pulmonary vessels and a layered segmentation of the lung volume to assess the diagnostic value of different vascular regions.

Materials and methods: We included 41 CPE, 41 acute pulmonary embolism (APE) and 41 normal controls (non-PE). 25 of the CPE patients had CTEPH confirmed by right heart catheterization. CNN classifiers were trained to identify CPE or CTEPH against a combined APE and non-PE group. Eleven masking schemes were applied for both classification tasks, resulting in 22 experiments. Model performances were compared using areas under the receiver operating characteristic curves (AUROC).

Results: The model achieved good performance in distinguishing CPE from non-PE and APE cases (cross-validation AUROC 0.80) using full lung volume MIPs, while performance decreased with reduced data. For CTEPH classification against non-PE and APE, the model reached AUROC 0.88 with full data and 0.86 using only the most proximal half of the lung volume, suggesting key diagnostic features reside centrally. Using an open-source segmentation model, which excludes proximal vessels, resulted in lower AUROCs (0.74 for CPE, 0.83 for CTEPH).

Conclusion: The cross-validation indicated that CPE and CTEPH could be identified from CTPA-derived MIP images, with performance improving as more vessels were included. The proximal vessels were most relevant for CTEPH detection.

Relevance statement: Our study shows that neural networks can identify chronic pulmonary embolism in CTPA and the role of different vascular regions in that task, with the potential to improve future imaging diagnostics in patients with chronic pulmonary embolism.

Key points: A convolutional neural network detects chronic thromboembolic pulmonary hypertension and chronic embolism from CTPA MIP projections. CTPA data were divided into four concentric anatomic layers for regional analysis. Central layers were most important for identifying CTEPH features. Network performance improved when more vessel regions were used as input.

Objective: Computed tomography-guided biopsies are needed to diagnose bone lesions, but can sometimes be challenging. We evaluated the feasibility and usefulness of photon-counting computed tomography (PCCT)-guided bone biopsies, focusing on real-time bone marrow oedema (BME) mapping to optimise diagnostic yield.

Materials and methods: This retrospective single-centre study included procedures performed from September 2024 to May 2025 using a first-generation dual-source PCCT scanner with Quantum HD mode. Ten consecutive patients underwent PCCT-guided bone biopsy with real-time BME reconstructions. The reference standard was established using histopathology or microbiological confirmation when available; clinical and ≥ 3-month radiologic follow-up for nondiagnostic or discordant results. Statistical analysis included descriptive statistics, independent unpaired t-tests, and correlation analysis (SPSS v22.0, RStudio).

Results: Ten patients, five women and five men, aged 60.5 ± 13.5 years (mean ± standard deviation), were included in the final analysis. The overall diagnostic yield was 70% (7/10), with a diagnostic accuracy of 87.5% (7/8) for cases with a definitive reference standard. Final diagnoses comprised tumour bone metastases (n = 7, 70%), bone osteomyelitis (n = 1, 10%), and bone marrow deposition disease (n = 2, 20%). Mean radiation dose (dose-length product) was 644.5 ± 112.1 mGy·cm. Monoenergetic 70-keV imaging showed significant differences between mean HU values of lytic (42.6) and sclerotic lesions (476.2) (p = 0.009), with a strong negative correlation between lesion morphology (sclerotic versus lytic) and monoenergetic 70-keV attenuation values (r = -0.84; p = 0.002).

Conclusion: PCCT-guided bone biopsy with real-time BME mapping proved feasible and showed encouraging diagnostic performance in this small exploratory cohort. Larger validation studies are needed.

Relevance statement: By combining monoenergetic images and BME mapping, PCCT-guided bone biopsy improves lesion visualisation, operator confidence, procedural efficiency, and overall safety for diagnostic tissue sampling and active disease targeting.

Key points: Accurate targeting of active disease within complex bone lesions during CT-guided biopsy remains challenging sometimes. PCCT-guided biopsy with real-time BME mapping can enhance lesion targeting and procedural efficiency. PCCT-guided biopsy may improve safety, diagnostic accuracy, and operator confidence.

Objective: Using radiomics to compute quantitative imaging features may reveal information beyond standard magnetic resonance imaging (MRI) metrics. We aim to investigate the test-retest repeatability of ¹⁹F MRI radiomic features in phantoms containing two perfluorocarbons and to validate these findings in a pilot in vivo mouse tumor model.

Materials and methods: Two phantoms containing perfluoropolyether (PFPE) or perfluoro-15-crown-5 ether (PFCE) were repeatedly scanned (intrasession and intersession) using a 7-T system equipped with a dual-tuned ¹H/¹⁹F volume coil. Radiomic features were extracted and assessed for stability using the concordance correlation coefficient (CCC) ≥ 0.85 and normalized dynamic range ≥ 0.90. A separate in vivo test-retest experiment was conducted in tumor-bearing mice injected with a PFPE nanoemulsion.

Results: A total of 194 scans and 772 segments were evaluated across the PFPE phantom, PFCE phantom, and in vivo experiments. In both phantoms, radiomic features displayed high intrasession repeatability (median CCC up to 0.886) but decreased intersession repeatability (median CCC down to 0.683). Intensity features were consistently more repeatable (p < 0.003) than shape or texture features. We found that 23.1% (466/2,013) of features were repeatable across phantoms. In vivo pilot scans showed that 86.1% (401/466) of these phantom-stable features, or ~20.0% overall, remained repeatable under physiological conditions.

Conclusion: Several ¹⁹F MRI-derived features exhibited excellent short-term repeatability, and a considerable proportion proved robust to intersession variability. These robust features may reliably capture ¹⁹F signals under both phantom and physiological conditions, paving the way for more quantitative imaging analysis in this modality and encouraging general reproducibility of data.

Relevance statement: KEY POINTS: We analyzed 194 ¹⁹F MRI scans and 772 segments obtained in phantoms at 7 T. Cross-agent stability identified 466 radiomic features meeting concordance correlation coefficient ≥ 0.85 and normalized dynamic range ≥ 0.90. Of these phantom-stable features, 401 of 466 remained stable in vivo in a tumor mouse model. Intensity features were most repeatable, while shape features were least stable across sessions. Median concordance correlation coefficient dropped from 0.886 intrasession to 0.683 intersession.

Objective: We used deep learning to generate synthetic, resembling in appearance, iodine-enhanced, mammograms from low-energy contrast-enhanced mammography (CEM) images.

Materials and methods: We retrospectively selected 140 CEM examinations. We trained a two-dimensional cycle-generative adversarial network on 390 images in 100 patients (195 breasts; 102 positive and 93 negative for lesion detection) using paired low-energy and iodine-enhanced images as input and output, respectively. We validated our model in 40 test patients (63 breasts; 37 positive and 26 negative for lesion detection) by calculating the contrast-to-noise ratio (CNR) for low-energy, synthetic, and clinical iodine-enhanced images and the mean absolute error (MAE) and similarity index metric (SSIM) between clinical and synthetic iodine-enhanced images regarding their changes from low-energy. Three radiologists scored (a-to-d) the test set images for background parenchymal enhancement (BPE) and lesion detection (yes/no) on clinical and synthetic images. The presence of artifacts was reported on all images.

Results: We observed a high correlation between clinical and synthetic iodine-enhanced images regarding their changes from low-energy: MAE, r = 0.99; SSIM, r = 0.80. CNR was -0.015/-0.16 ± 0.23/0.05 (mean ± standard deviation) for clinical/synthetic, respectively. A "halo" artifact present in above 50% of the clinical iodine-enhanced images was corrected in the synthetic ones. On synthetic images, BPE (scores a-b versus c-d) was 85.8% accurate. Lesion detection accuracy was 89.4% and 79.4%, sensitivity 87.4 and 72.1%, and specificity 92.3% and 90.0% for clinical and synthetic images, respectively.

Conclusions: Deep learning holds the potential to generate synthetic iodine-enhanced mammograms from low-energy images.

Relevance statement: Radiologists could perform some clinical tasks, such as lesion detection and BPE estimation on synthetic iodine-enhanced images, without contrast injection.

Key points: Our deep learning model generated synthetic iodine-enhanced images that visually resembled the clinical iodine-enhanced images. Radiologists could use the synthetic images to perform clinical tasks, such as lesion detection and BPE evaluation. Our model can improve image quality by removing common artifacts, including the breast-in-breast (halo). Our method is a way to combine the benefits of CEM while sparing the need for contrast media.

Objective: Collateral circulation is a key determinant of functional outcome after large vessel occlusion (LVO) and informs thrombectomy decisions. However, collateral grading is rater-dependent and error-prone. We developed an automated cerebrovascular radiomics pipeline to establish objective collateral scoring on computed tomography angiography (CTA).

Materials and methods: We retrospectively analyzed admission CTAs from 343 LVO patients in the MR CLEAN trial, split into training/validation (n = 274) and testing (n = 69) sets. Vessel regions of interest were segmented using nnU-Net models trained on 40 arterial tree CTAs and 125 multiclass circle of Willis (CoW) cases. Radiomics features were extracted from vascular regions. Predictive features were identified, and a random forest classifier was trained to distinguish sufficient (> 50%) from insufficient (≤ 50%) collateral status according to the Tan score system. Performance was compared to the atlas-based middle cerebral artery (MCA) mask model and validated on an external cohort of 140 acute LVO patients.

Results: Segmentation models accurately annotated cerebral arteries (95th percentile Hausdorff distance 4.49, Dice similarity coefficient 0.83) and CoW segments (2.27 and 0.81, respectively). After feature selection, 6 top features were identified for vessel-tree radiomics, 98 for MCA mask-based radiomics, and 32 for a combined vessel-tree/CoW model. Vessel-tree outperformed MCA mask model on both internal (area under the receiver operating characteristic curve (AUROC): 0.88 versus 0.82) and external (AUROC: 0.83 versus 0.66) test sets. Adding CoW features further improved performance, achieving 0.87 AUROC.

Conclusion: We presented a fully automated generalizable CTA radiomics approach for objective collateral scoring in acute LVO.

Relevance statement: This study introduces a fully automated CTA cerebrovascular radiomics pipeline that objectively assesses collateral status in patients with acute ischemic stroke. Combining vessel-tree and circle of Willis features improved collateral score prediction accuracy and generalizability, supporting more reliable, data-driven decision-making in acute large vessel occlusion management.

Key points: Collateral circulation status informs prognosis and guides treatment in acute stroke, but grading is rater-dependent; our pipeline standardizes collateral assessment. We propose a CTA radiomics approach, trained and validated on multicenter data, externally tested on an independent cohort, demonstrating high effectiveness and generalizability. Automated and reliable collateral scoring has the potential to reduce inter-rater variability, improve workflow efficiency, and support individualized treatment decisions.

Objective: Radiogenomics promises noninvasive tumor profiling; however, the extent to which imaging morphology reflects tumor lineage versus host-organ milieu remains unclear. This study aimed to quantify the relative influence of tumor type and anatomical environment on contrast-enhanced computed tomography (CT) radiomic phenotypes.

Materials and methods: A discovery cohort of 1,598 patients (10,485 lesions) and an external validation cohort of 2,440 patients (6,597 lesions) underwent portal-venous-phase CT. After manual segmentation, lesion-level radiomic features were standardized and embedded using t-distributed stochastic neighbor embedding. Bayesian-optimized agglomerative clustering defined morphology-based groups. Concordance with the primary tumor site (lineage) and anatomical environment was quantified using bootstrapped adjusted Rand indices (ARI); the silhouette score assessed clustering quality. Feature-class (shape, intensity, texture) and mask-erosion experiments probed mechanistic drivers.

Results: Six morphological clusters were identified in the discovery set (silhouette = 0.44). Morphology aligned more strongly with environment (mean ARI = 0.37) but poorly with lineage (mean ARI = 0.04; p < 0.010); this pattern held externally. In solid organ metastases, environment dominance was even stronger (mean ARI = 0.60 versus 0.05; p < 0.010). Intensity and texture drove the morphological association with anatomical environment (ARI = 0.64-0.56) more than shape (ARI = 0.06). When the periphery of the tumor was eroded, the same patterns were observed, implicating the tumor core.

Conclusion: Across organs and tumor types, tumor morphological phenotype on CT imaging is largely driven by a host tissue-related environmental "imprint" rather than the primary tumor site.

Relevance statement: Context-aware modeling is essential for reliable radiomic biomarkers and could motivate a two-step AI pipeline that first identifies the organ habitat and refines lineage-specific predictions.

Key points: In a large, multicenter cohort, tumors exhibited distinct morphological clustering. These clusters did not align with primary tumor sites (ARI = 0.04). Stronger associations emerged between morphological clusters and the local anatomical environment (ARI = 0.37). Stratification by lesion type revealed even stronger associations between local anatomical context and solid organ metastases (ARI = 0.60).

Objectives: We compared three customized nnU-Net models (A: baseline two-dimensional (2D); B: 2D + region-growing; C: three-dimensional (3D) + region-growing) for automated detection and blood clot volume (BCV) quantification of acute pulmonary embolism (PE) on computed tomography pulmonary angiography (CTPA), and to explore the association between BCV and clinical outcome.

Materials and methods: We retrospectively screened 9,715 CTPA examinations (2015‒2024) to develop a dataset of 874 PE-positive and 339 PE-negative cases. A stratified subset (n = 437) with manually refined ground-truth segmentations was used for model training and internal validation. Region-growing in Models B and C included a 5-voxel negative buffer. Internal testing was performed on 776 cases (Humanitas dataset). External testing was performed on the public RSPECT-RSNA dataset. Performance metrics included accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) at zero-clot and for optimized BCV threshold. Correlations between BCV, survival, and major adverse cardiovascular events (MACE) were analyzed.

Results: Model C achieved the highest AUROC on external testing (0.868), outperforming Model A (0.843) and Model B (0.846). On internal testing at ROC-optimized threshold, Model C showed the highest accuracy (85.5%) and AUROC (0.909) compared to Model A (73.4%, 0.784) and Model B (76.0%, 0.816). Model C achieved 83.6% sensitivity and 79.5% accuracy at the zero-clot threshold on external data. BCV was not significantly associated with MACE or survival (p = 0.600).

Conclusion: A locally trained 3D nnU-Net with region-growing demonstrated superior performance and generalizability on external data for automated PE detection on CTPA. However, BCV was not predictive of short-term clinical outcomes.

Relevance statement: A locally developed nnU-Net models integrating volumetric 3D segmentation with region-growing offer robust, clinically acceptable performance for the detection of acute pulmonary embolism without the need for ROC-based thresholds.

Key points: Our 3D nnU-Net model automates clot detection on CT scans in seconds and shows numerically higher performance than the 2D models. Built on local data, this framework enables institution-specific model training and validation to complement European conformity‒CE-marked tools and assess performance locally. High-sensitivity volumetric quantification reduces missed emboli, paving the way for personalized risk stratification and improved patient outcomes.

Objective: Pulmonary arterial hypertension is a severe complication of systemic lupus erythematosus (SLE). Current screening methods often miss early vascular changes. This study aimed to characterize subclinical pulmonary hemodynamic alterations in SLE patients without known pulmonary arterial hypertension using four-dimensional (4D) flow cardiovascular magnetic resonance (CMR) and to investigate their association with left ventricular diastolic function.

Materials and methods: Twenty-five SLE patients without known pulmonary arterial hypertension and 25 age-matched healthy controls were enrolled. All participants underwent 3-T 4D flow CMR to quantify hemodynamic parameters, including wall shear stress (WSS), flow volume, and relative pressure in the pulmonary arteries. SLE patients were further stratified based on echocardiographic assessment of diastolic function to analyze hemodynamic coupling.

Results: Compared to controls, SLE patients exhibited significantly lower maximum WSS in the main pulmonary artery (0.29 versus 0.33 Pa, p = 0.040) and asymmetric flow redistribution, characterized by higher relative pressure in the left pulmonary artery (0.54 versus 0.30 mmHg, p = 0.008) and increased flow rate in the right pulmonary artery (3.51 versus 2.90 L/min, p = 0.015). Qualitative analysis revealed vortical flow patterns in SLE patients. Subgroup analysis demonstrated that the reduction in WSS was primarily driven by patients with diastolic dysfunction (p = 0.006 versus controls).

Conclusion: SLE patients without pulmonary arterial hypertension exhibit distinct subclinical pulmonary hemodynamic alterations, including lower WSS and flow asymmetry. These alterations are intimately coupled with left ventricular diastolic dysfunction, suggesting that 4D flow CMR serves as a sensitive noninvasive tool for early risk stratification in this population.

Relevance statement: 4D flow CMR identifies subclinical pulmonary hemodynamic alterations coupled with diastolic dysfunction in SLE patients, serving as a sensitive noninvasive tool for early risk stratification before irreversible vascular remodeling occurs.

Key points: SLE patients without known pulmonary arterial hypertension show early pulmonary blood flow changes. 4D flow CMR detected asymmetric pulmonary flow redistribution in SLE patients. SLE patients exhibited altered left atrial function despite normal ventricles. Pulmonary flow changes correlated with left atrial remodeling in SLE. 4D flow CMR detects subclinical pulmonary hemodynamic differences in SLE.