Radiology-Artificial Intelligence最新文献

Purpose To evaluate random convolutions as an augmentation strategy for improving domain generalization of deep learning-based segmentation models in medical imaging. Materials and Methods In this retrospective study, a random convolution-based augmentation strategy was applied to abdominal organ segmentation (AbdomenCT-1k dataset: 361 CT images; Abdominal Multi Organ Segmentation [AMOS] dataset: 298 CT and 59 MRI scans) and brain tissue segmentation (Information eXtraction from Images [IXI] dataset: 504 T1-weighted images from Guy's and Hammersmith Hospitals, 146 paired T1-weighted and T2-weighted images from the Institute of Psychiatry). Performance was compared with baseline and state-of-the-art segmentation models (TotalSegmentator and deepAtropos). Random convolution configurations were analyzed for effects on in- and out-of-domain performance. Results The random convolution-enhanced U-Net achieved in-domain Dice scores comparable to state-of-the-art baselines (CT: 0.93 vs TotalSegmentator: 0.95; T1-weighted imaging: 0.83 vs deepAtropos: 0.79). Out-of-domain Dice scores were significantly higher (MRI: 0.93, T2-weighted imaging: 0.52) compared with baselines (TotalSegmentator in MRI: 0.85, deepAtropos in T2-weighted imaging: 0.33; false discovery rate-adjusted P < .001). Augmentation probability and configuration influenced the trade-off between in- and out-of-domain performance. Conclusion Random convolutions yielding more robust segmentation models generalized better to unseen domains than models trained without random convolutions and are compatible with diverse segmentation architectures. Keywords: MR-Imaging, CT, Supervised Learning, Segmentation, Abdomen/GI, Experimental Investigations Supplemental material is available for this article. © RSNA, 2025 See also commentary by Mathai in this issue.

Purpose To evaluate the characteristics of breast cancers detected and missed by artificial intelligence-based computer-assisted diagnosis (AI-CAD) during screening mammography. Materials and Methods This retrospective secondary analysis was conducted using data from the Artificial Intelligence for Breast Cancer Screening in Mammography trial (ClinicalTrials.gov: NCT05024591), a prospective, multicenter cohort study performed from 2021 to 2022. AI-CAD results were categorized into nine subgroups based on abnormality scores (in 10% increments). Positive predictive values of recall (PPV1s) were calculated for each subgroup and by breast density, and AI-CAD scores were compared with mammographic and pathologic features. Results A total of 24 543 women (mean age ± SD, 59.8 years ± 11.2), including two with bilateral cancer, were included; 148 cancers were confirmed by pathologic evaluation after 1 year of follow-up. AI-CAD results were negative in 23 010 cases (93.8%) and positive in 1535 (6.2%). The overall PPV1 was 8.7% (133 of 1535), with a sensitivity of 89.9% and specificity of 94.3%; PPV1 increased with higher abnormality scores but remained below 3% in groups 1 and 3 for dense breasts. AI-CAD detected 3.4% (five of 148) of cancers missed by radiologists but missed 8.1% (12 of 148) that were detected at radiologist recall. Abnormality scores were lower in patients presenting with mammographic asymmetry (P = .001) and luminal A subtype (P = .032). Conclusion AI-CAD shows potential to improve breast cancer detection in screening programs and to support radiologists in mammogram interpretation. Understanding the imaging and pathologic features of cancers detected or missed by AI-CAD may enhance its effective clinical application. Keywords: Breast Cancer, Mammography, AI CAD Clinical trial registration no. NCT05024591 © RSNA, 2025 See also commentary by Do and Bahl in this issue.

Purpose To develop and evaluate a deep learning-based brain extraction model, CTA-BET, capable of providing accurate brain segmentation for CT angiography (CTA) and non-contrast-enhanced CT (NCCT) images. Materials and Methods In this retrospective study, CTA-BET was trained using CTA data from multi-institutional cohorts (n = 100 patients) and validated on an external CTA dataset (n = 50 patients). NCCT validation was performed using the publicly available CQ500 dataset (n = 132 patients). The model's performance was compared with five benchmark noncommercial brain extraction tools. Dice score, Hausdorff distance, and z score-normalized histograms were used to evaluate segmentation performance. Results The CTA-BET model outperformed all benchmark models, achieving a mean Dice score of 0.99 (95% CI: 0.99, 0.99) on CTA data (P < .001 for all comparisons) and 0.98 (95% CI: 0.98, 0.99) on NCCT images (P < .001 for all comparisons). In terms of Hausdorff distance, CTA-BET demonstrated higher performance compared with other benchmark tools on CTA images (P < .001 for all comparisons). Conclusion CTA-BET outperformed benchmark brain extraction tools on both CTA and NCCT images, providing a robust and accurate solution that could enhance automated imaging analysis in clinical and research settings. Keywords: CT-Angiography, Head/Neck, Brain/Brain Stem, Computer Applications-3D, Comparative Studies, Experimental Investigations, Technology Assessment, Segmentation, Convolutional Neural Network (CNN) Supplemental material is available for this article. © RSNA 2025.

Metadata, which refers to nonimage information such as patient identifiers, acquisition parameters, and institutional details, have long been the primary focus of de-identification efforts when constructing datasets for artificial intelligence applications in medical imaging. However, it is now evident that information intrinsic to the image itself, at the pixel level (eg, intensity values), can also be exploited by deep learning models, potentially revealing sensitive patient data and posing privacy risks. This report discusses both metadata and sources of identifiable information in medical imaging studies, highlighting the potential risks of overlooking their presence. Privacy-preserving approaches such as federated learning and synthetic data generation are also reviewed, with emphasis on their limitations-particularly vulnerabilities to model inversion and inference attacks-that must be considered when developing and deploying artificial intelligence in medical imaging. Keywords: Privacy, Metadata, Synthetic, Federated Learning, Anonymization De-identification ©RSNA, 2025.

Purpose To model the distribution of CT-derived whole-body anatomic volumes across adulthood and establish comprehensive cross-sectional and longitudinal reference charts, addressing the current lack of non-brain CT-based whole-body standards. Materials and Methods Retrospective CT scans acquired from March 2017 to April 2025 (189,710 scans, 106,563 patients) from the institutional PACS and two external datasets (19,393 and 1,158 patients, respectively), were automatically segmented into 104 structures (totaling 7.8 million volumes). An automated quality control pipeline, incorporating a novel outlier removal strategy based on strong correlation between organ sizes, ensured data reliability. Cross-sectional normative models were constructed using Generalized Additive Models for Location, Scale, and Shape (GAMLSS) to capture non-linear age effects through fractional polynomial functions. A Generalized Additive Mixed Model (GAMM) was employed for longitudinal analyses to assess within-subject changes over follow-up visits. Results All anatomic structures followed complex, non-linear age trajectories, with marked sex differences and distinct CT contrast effects on vascular structures. Bootstrap resampling confirmed the stability and precision of these volume trajectories in both central tendency and variability. An exemplary cardiomegaly case-control analysis showed significantly increased centile scores (P < .001) for heart volume. The longitudinal analysis further revealed significant age-sex interactions influencing within-subject trajectories. Conclusion Cross-sectional and longitudinal reference models were developed from CT-derived anatomic volumes that map the trajectories of body structural change across adulthood. These body charts facilitate robust quantification of individual deviations via centile scores. ©RSNA, 2025.

Purpose To evaluate the association between radiologist and stand-alone AI concordance in screening mammography interpretations and future breast cancer risk and to assess how breast density influences this relationship. Materials and Methods This retrospective study included Korean women who underwent digital screening mammography between January 2009 and December 2018. Incidental breast cancers, defined as those diagnosed within 1 year of screening, were excluded using the National Cancer Registry data. A commercial AI system retrospectively analyzed mammograms, with a 10% malignancy probability threshold. Patients were categorized into four groups based on AI-radiologist concordance: Concordant-Negative, Radiologist-Positive/AI-Negative, Radiologist-Negative/AI-Positive, and Concordant-Positive. Breast density was categorized using Breast Imaging Reporting and Data System classification. Cox proportional hazards models estimated adjusted hazard ratios (HRs) and 95% CIs. Results Over a median 7.3-year follow-up, 1011 breast cancers occurred among 82 899 women (mean age, 43.4 years ± 8.6 [SD]). Breast cancer incidence was the highest in the Concordant-Positive group (5-year cumulative incidence, 37.4 per 1000 person-years) and the lowest in the Concordant-Negative group (5.9 per 1000 person-years). Compared with the Concordant-Negative group, adjusted HRs for incidental breast cancer were 2.30 (P < .001) for the Radiologist-Negative/AI-Positive, 1.15 (P = .17) for the Radiologist-Positive/AI-Negative, and 4.51 (P < .001) for the Concordant-Positive groups. Risk increase in the Concordant-Positive group was consistent across breast densities; in dense breasts, elevated risk occurred only with positive AI. Conclusion Positive screening mammography findings identified by both radiologists and stand-alone AI (Concordant-Positive group) were associated with he highest 5-year breast cancer incidence of 3.74%, exceeding the 3% threshold for considering chemoprevention or supplemental imaging. Keywords: Mammography, Breast Neoplasms, Artificial Intelligence, Screening, Concordance Supplemental material is available for this article. © RSNA, 2025.

Purpose To assess the effectiveness of an explainable deep learning model, developed using multiparametric MRI features, in improving diagnostic accuracy and efficiency of radiologists for classification of focal liver lesions (FLLs). Materials and Methods FLLs 1 cm or larger in diameter at multiparametric MRI were included in the study. The nn-Unet and Liver Imaging Feature Transformer models were developed using retrospective data from the Ruijin Hospital (January 2018-August 2023). The nnU-Net was used for lesion segmentation and the Liver Imaging Feature Transformer model for FLL classification. External testing was performed on data from the Xinjiang Production and Construction Corps Hospital, the First Affiliated Hospital of Soochow University, and Xinrui Hospital (January 2018-December 2023), with a prospective test set obtained from January to April 2024. Model performance was compared with radiologists, and impact of model assistance on junior and senior radiologist performance was assessed. Evaluation metrics included the Dice similarity coefficient and accuracy. Results A total of 2131 individuals with FLLs (mean age, 56 years ± 12 [SD]; 1476 female patients) were included in the training, internal test, external test, and prospective test sets. Average Dice similarity coefficient values for liver and tumor segmentation across the three test sets were 0.98 and 0.96, respectively. Average accuracy for features and lesion classification across the three test sets were 93% and 97%, respectively. Readings assisted by the Liver Imaging Feature Transformer model improved diagnostic accuracy (average 5.3% increase, P < .001), reduced reading time (average 34.5 seconds decrease, P < .001), and enhanced confidence (average 0.3-point increase, P < .001) of junior radiologists. Conclusion The proposed deep learning model accurately detected and classified FLLs, improving diagnostic accuracy and efficiency of junior radiologists. Keywords: Liver, MR-Dynamic Contrast Enhanced, Convolutional Neural Network (CNN), Deep Learning Algorithms, Machine Learning Algorithms, Feature Detection, Vision, Application Domain Supplemental material is available for this article. © RSNA, 2025 See also commentary by Adams and Bressem in this issue.

Purpose To introduce an open-source deep learning brain segmentation model for fully automated brain MRI segmentation, enabling rapid segmentation and facilitating large-scale neuroimaging research. Materials and Methods In this retrospective study, a deep learning model was developed using a diverse training dataset of 1900 MRI scans (patients aged 24-93 years, with a mean of 65 years ± 11.5 [SD]; 1007 female, 893 male) with reference labels generated using a multi-atlas segmentation method with human supervision. The final model was validated using 71 391 scans from 14 studies. Segmentation quality was assessed using Dice similarity and Pearson correlation coefficients with reference segmentations. Downstream predictive performance for brain age and Alzheimer disease was evaluated by fitting machine learning models. Statistical significance was assessed using Mann-Whitney U and McNemar tests. Results The DLMUSE model achieved high correlation (r = 0.93-0.95) and agreement (median Dice scores, 0.84-0.89) with reference segmentations across the testing dataset. Prediction of brain age using DLMUSE features achieved a mean absolute error of 5.08 years, similar to that of the reference method (5.15 years, P = .56). Classification of Alzheimer disease using DLMUSE features achieved an accuracy of 89% and F1 score of 0.80, which were comparable to values achieved by the reference method (89% and 0.79, respectively). DLMUSE segmentation speed was over 10 000 times faster than that of the reference method (3.5 seconds vs 14 hours). Conclusion DLMUSE enabled rapid brain MRI segmentation, with performance comparable to that of state-of-the-art methods across diverse datasets. The resulting open-source tools and user-friendly web interface can facilitate large-scale neuroimaging research and wide utilization of advanced segmentation methods. Keywords: Convolutional Neural Network (CNN), Deep Learning Algorithms, Machine Learning Algorithms, Segmentation, Application Domain, Supervised Learning, MRI, Brain/Brain Stem Supplemental material is available for this article. © RSNA, 2025.