Radiology. Imaging cancer最新文献

Purpose To develop a fully automated hybrid approach to predict sacral tumor types from preoperative noncontrast CT (NCCT ) images. Materials and Methods In this retrospective, multicenter study, scans were available in 690 patients who had histopathologically confirmed preoperative sacral NCCT performed between January 2011 and May 2024. A fully automated hybrid model integrated two deep convolutional neural network models (model 1 and model 2) through a fully automated pipeline. Model 1 segments tumors and hip bones automatically from NCCT images, producing masks that are used by model 2. For the first time, the hip bone was used as a reference frame for tumor localization. The second model, CL-MedImageNet, is an innovative six-classification model that allows the simultaneous input of tumor images, clinical data, and location information. This streamlined, automated system ensures efficient data integration and processing between the two models. The efficacy of the model was assessed in comparison to that of radiologists, using metrics including the area under the curve (AUC), F1 score, and confusion matrix. Results In all, 690 patients (mean age, 46 years ± 17 [SD]; 377 male patients) were included. Segmentation achieved mean Dice coefficients of 0.82 ± 0.11 (validation), 0.81 ± 0.12 (internal test), and 0.81 ± 0.12 (external test) after postprocessing; interobserver Dice coefficient was 0.96. The CL-MedImageNet classifier attained macro average AUCs of 0.89 (95% CI: 0.83, 0.93), 0.88 (95% CI: 0.84, 0.92), and 0.87 (95% CI: 0.79, 0.92) in validation, internal, and external test sets, respectively, with macro average F1 scores of 0.63, 0.63, and 0.56. The highest achieved precision and sensitivity were both 0.66 across all sets. CL-MedImageNet outperformed radiologists (macro average AUCs, 0.87 vs 0.80, P = .002; 0.87 vs 0.83, P = .45). Conclusion The fully automated NCCT-based CL-MedImageNet pipeline demonstrated high segmentation accuracy and robust six-class classification, outperforming expert radiologists. Keywords: Applications - CT, Deep Learning, Radiomics, Segmentation, Skeletal-Axial, Pelvis, Sacral Tumors Supplemental material is available for this article. © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license.

Purpose To develop a radiomics model based on hepatobiliary phase gadolinium ethoxybenzyl-diethylenetriaminepentaacetic acid (EOB)-enhanced MRI features at the tumor margin to predict microvascular invasion in high-risk solitary hepatocellular carcinoma (HR-sHCC), determine the optimal margin region, and explore the underlying biologic mechanisms. Materials and Methods This retrospective study included patients with HR-sHCC from three medical centers between April 2015 and December 2022. Radiomics features were extracted from 121 volumes of interest (VOIs) at the tumor margin at EOB MRI. Nine combinations of statistical and machine learning methods were used to construct and validate the optimal margin region-based radiomics model. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and patient stratification was evaluated with Kaplan-Meier and log-rank analyses. RNA sequencing data underwent differential expression analysis with DESeq2, followed by Kyoto Encyclopedia of Genes and Genomes (ie, KEGG) and Gene Ontology (ie, GO) enrichment, and immune cell infiltration was assessed using xCell and EPIC. Results A total of 436 patients (mean age, 57.7 years ± 8.8 [SD]; 352 male) were included: 254 in the training, 108 in the internal test, and 74 in the external test cohorts. Receiver operating characteristic analysis showed AUCs of 0.80 (95% CI: 0.74, 0.86), 0.76 (95% CI: 0.66, 0.85), and 0.72 (95% CI: 0.58, 0.86), respectively. The model effectively stratified patients by overall and disease-free survival (all P < .05). RNA sequencing revealed extracellular matrix remodeling, transforming growth factor-β signaling, and M2 macrophage infiltration in high optimal margin region-score tumors. Conclusion The optimal margin region-based radiomics model, derived from EOB MRI, effectively captured tumor margin heterogeneity. Keywords: MRI, Machine Learning, Radiomics, Radiogenomics, Abdomen/GI, Liver, Surgery, High-Risk Solitary Hepatocellular Carcinoma, Tumor Margin, Microvascular Invasion, Gd-EOB-DTPA-enhanced MRI, OATP1B3 © The Author(s) 2026. Published by the Radiological Society of North America under a CC BY 4.0 license. Supplemental material is available for this article.

Purpose To assess therapeutic and prognostic implications of perfusion characterization by fractal analysis using routine MRI in high-risk primary neuroblastomas and to establish a pathophysiologic connection between vascularity phenotype, perfusion imaging characteristics, and treatment response. Materials and Methods In a retrospective cohort study across 30 centers, MRI data of patients with high-risk neuroblastoma (June 2005-February 2021) were collected at the time point of diagnosis (TP1) and after induction chemotherapy before surgery (TP2), with data split into separate discovery (single-center) and validation cohorts (29 centers). Fractal analysis was performed on contrast-enhanced, fat-saturated, T1-weighted sequences at both time points to obtain voxel-wise local fractal dimension (FD) maps for predicting volumetric tumor response. The association of global FD with event-free survival (EFS) was assessed using a Cox proportional hazards model. Additionally, FD was calculated from CD34-stained endothelium in selected histologic tumor samples. Accuracy of response prediction, prognostic value for EFS, and correlation between FD of immunohistochemical vascularity and MRI-derived perfusion were also evaluated. Results In 73 patients (median age, 3 years [IQR, 3]; 39 male patients; discovery cohort, n = 36; validation cohort, n = 37), local FD maps helped predict volumetric tumor response to induction chemotherapy between TP1 and TP2 with good accuracy (root mean squared error, 47.78 mL; R² = 0.94; P < .001), visualizing intratumor high perfusion complexity in areas with low response potential. In multivariate Cox proportional hazards modeling, MYCN status (hazard ratio, 2.30; 95% CI: 1.16, 4.55; P = .017) and global FD at TP2 (hazard ratio, 0.65; 95% CI: 0.47, 0.88; P = .006) were significantly associated with EFS. Complexity of both CD34-immunohistochemical microvascularity (1.23 ± 0.09 [SD] to 1.44 ± 0.07, P < .001) and MRI perfusion (3.40 ± 0.04 to 3.53 ± 0.07, P < .001) increased throughout induction chemotherapy. Conclusion Fractal analysis of MRI-derived perfusion complexity was associated with spatial heterogeneity of chemotherapy response and stratified prognosis in MYCN nonamplified high-risk neuroblastoma, supporting its potential as an imaging biomarker linked to microvascular architecture. German Clinical Trial Registry: DRKS00023442 Keywords: Pediatrics, MR-Imaging, Nervous-Peripheral, Fractal Analysis, Tissue Characterization, Tumor Response Supplemental material is available for this article. © RSNA, 2025.

Purpose To evaluate the intermodality and interobserver agreement of fast field echo resembling a CT using restricted echo-spacing (FRACTURE) MRI with conventional radiography and CT in assessing bone tumors. Materials and Methods This retrospective study included patients with histopathologically confirmed bone tumors who underwent FRACTURE MRI and radiography over a 10-month period; a subset also underwent CT. Two independent readers assessed lesion features across modalities, with agreement calculated using Cohen κ. Results Among 161 patients (median age, 18 years [IQR, 13.5-31]; 95 male), FRACTURE showed near-perfect/perfect agreement with conventional radiography for lesion type (κ = 0.96 and 0.86, for reader 1 and reader 2, respectively) and treatment response (κ = 1.00, 1.00), near-perfect/substantial agreement for both fractures (κ = 0.95, 0.76) and lesion location (κ = 0.95, 0.79), and near-perfect/moderate agreement for matrix content (κ = 0.89, 0.56). Agreement was substantial for extraosseous soft tissue (κ = 0.74, 0.65) and osteochondral defects (κ = 0.75, 0.75) and moderate for lesion margins (κ = 0.42, 0.59). Cortical integrity showed moderate to near-perfect agreement (range, κ = 0.46-0.85) and periosteal reactions fair to near-perfect agreement (range, κ = 0.34-0.91), depending on subtype and reader. In the CT subset (n = 69), FRACTURE maintained near-perfect/perfect agreement for lesion type (κ = 0.90, 0.84), location (κ = 1.00, 0.94), extraosseous soft tissue (κ = 0.96, 0.85), fractures (κ = 0.95, 0.96), treatment response (κ = 1.00, 1.00), and osteochondral defects (κ = 1.00, 1.00). Near-perfect/substantial agreement was present for matrix content (κ = 0.81, 0.61). Cortical integrity and periosteal reaction showed fair to near-perfect agreement (range, κ = 0.25-0.81, 0.33-0.91). Conclusion FRACTURE MRI demonstrated high agreement with radiography and CT in evaluating most bone tumor features. Keywords: Conventional Radiography, MR Imaging, Skeletal-Appendicular, FRACTURE Sequence, 3D T1 GRE, Bone Tumors, MRI, Osteosarcoma Supplemental material is available for this article. © RSNA, 2026.

Purpose To develop and evaluate an image-to-image conditional generative adversarial network (cGAN) for translating dynamic contrast-enhanced (DCE) MRI data to vascular pharmacokinetic permeability maps. Materials and Methods Retrospective breast cancer DCE MR images from The Cancer Imaging Archive acquired between April 1996 and January 1998 were used to assess the developed cGAN. The extended Tofts model (ETM) was applied to establish reference standard volume transfer constant (K^trans) maps. The cGAN was trained to learn relationships between DCE MR data and ETM K^trans maps. Linear regression was applied to determine agreement between the ETM and cGAN. Logistic regression and paired t tests were used to assess predictive capabilities of pathologic response. Results Twenty DCE MRI scans (n = 2400 sections) from 10 female patients (mean age, 45 years ± 12 [SD]) were analyzed. Computation time was reduced over 1000-fold using the cGAN compared with the ETM. The cGAN K^trans maps exhibited excellent spatial agreement and high structural similarity to the ETM, with low errors (normalized root mean squared error ≤0.32; normalized mean absolute error ≤0.16) and a strong correlation (R² ≥ 0.98). Patients with pathologic complete response demonstrated a 60% reduction in cGAN K^trans (P = .01) after the first cycle of neoadjuvant chemotherapy, closely matching ETM K^trans (59%, P = .02). In contrast, patients without pathologic complete response showed a modest reduction in cGAN K^trans (17%, P = .13), still in good agreement with the ETM (15%, P = .19). Percentage of K^trans change effectively distinguished patients with or without pathologic complete response (C statistic = 1.0) for both models. Conclusion The DCE to pharmacokinetic cGAN offers promise for standardizing pharmacokinetic analysis and reducing computational complexity at DCE MRI. Moreover, this approach demonstrated potential for early prediction of breast cancer responses to neoadjuvant chemotherapy. Keywords: Dynamic Contrast-enhanced MRI, Vascular Permeability, Image-to-Image Conditional Generative Adversarial Network, Breast Cancer, Neoadjuvant Chemotherapy © RSNA, 2025.

Purpose To evaluate the performance of a CT-based model combining two-dimensional (2D) and two-and-a-half-dimensional (2.5D) deep learning (DL) with radiomic features in predicting neoadjuvant chemoimmunotherapy response in patients with esophageal squamous cell carcinoma (ESCC). Materials and Methods In this retrospective study, patients with ESCC from Sun Yat-sen Cancer Center between May 2020 and January 2023 were divided into training (80%) and internal validation (20%) groups, while an external testing group was obtained from Nanfang Hospital between January 2021 and March 2023. Radiomic features were extracted manually, while 2D and 2.5D deep transfer learning (DTL) features were derived from pretrained DL networks. The optimal model was selected based on a comparison of the areas under the receiver operating characteristic curves (AUCs). Results In total, 251 patients (mean age, 59.91 years ± 7.63; 209 male and 42 female) were included in the study, with 157 and 94 patients from centers 1 and 2, respectively. The support vector machine (SVM) model outperformed the other radiomic and DL models, while ResNet18 had the best predictive performance among the 2D and 2.5D DL models. The SVM model with ResNet18-based DTL features showed the best performance, achieving AUC values of 0.85 (95% CI: 0.76, 0.91) for 2D DTL and 0.84 (95% CI: 0.75, 0.91) for 2.5D DTL in the external testing group. Conclusion A fusion model integrating 2D and 2.5D DTL and radiomic features effectively predicted the neoadjuvant chemoimmunotherapy response in patients with ESCC. Keywords: Deep Learning, Artificial Intelligence, Prognosis & Prediction, Esophagus, CT Supplemental material is available for this article. © RSNA 2026.