Pub Date : 2026-03-01Epub Date: 2025-12-13DOI: 10.1016/j.acra.2025.11.039
Alina Carolin Seifert MD , Markus M. Obmann MD , Hanns-Christian Breit MD , Anastassia Korolenko RT , Omar Darwish PhD , Matthias Fenchel PhD , Daniel T. Boll MD , Jan Vosshenrich MD
Rationale and Objectives
Low-field diffusion-weighted imaging (DWI) is limited by inherently lower signal-to-noise ratios. This study evaluated the technical feasibility, image quality, and apparent diffusion coefficient (ADC) reliability of deep learning (DL)-enhanced DWI at low field strength.
Materials and Methods
Thirty-three healthy volunteers (mean age: 30 ± 4 years; 13 men) underwent 0.55 T abdominal magnetic resonance imaging from 07/2024-09/2024. Conventional and DL-enhanced DWI were acquired with two b values (b50 s/mm2, b800 s/mm2). Three fellowship-trained abdominal radiologists rated the scans for quality parameters and artifacts (Likert scales, 1-5). Quantitative b50 and b800 signal intensity and ADC values were measured in the right and left liver lobes, spleen and Th12 vertebral body by one reader. Interreader reliability, intermethod concordance, and Bland-Altman analyses were performed.
Results
DL-enhanced scans demonstrated higher quality across all diffusion weightings (all: P < .001): b50 images (median: 4 [IQR: 4-5] vs 3 [3-4]), b800 images (4 [4-4] vs 3 [3-3]), and ADC maps (4 [4-4] vs 3 [3-3]) with good or better interreader agreement (κ ≥ 0.63). Image noise, spatial resolution, organ sharpness and artifacts were rated better for DL-enhanced DWI (all: P < .001).
Mean b50 and b800 signal intensities were lower with DL reconstruction (all: P < .001). ADC values showed at least strong intermethod correlation (r ≥ 0.61; all: P < .001) with mean differences of 0.3%-4.5% in normal tissues. Bland-Altman plots and interchangeability analysis confirmed intermethod ADC value deviations within expected margins of ±20% (all: P < .001).
Conclusion
DL superresolution reconstruction enables faster and higher quality abdominal DWI at 0.55 T without relevant ADC deviations, supporting interchangeability with conventional DWI.
{"title":"Deep Learning-Enhanced Diffusion-Weighted Imaging of the Abdomen at 0.55 T: Image Quality and Apparent Diffusion Coefficient Calculation Interchangeability in Healthy Volunteers","authors":"Alina Carolin Seifert MD , Markus M. Obmann MD , Hanns-Christian Breit MD , Anastassia Korolenko RT , Omar Darwish PhD , Matthias Fenchel PhD , Daniel T. Boll MD , Jan Vosshenrich MD","doi":"10.1016/j.acra.2025.11.039","DOIUrl":"10.1016/j.acra.2025.11.039","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>Low-field diffusion-weighted imaging (DWI) is limited by inherently lower signal-to-noise ratios. This study evaluated the technical feasibility, image quality, and apparent diffusion coefficient (ADC) reliability of deep learning (DL)-enhanced DWI at low field strength.</div></div><div><h3>Materials and Methods</h3><div>Thirty-three healthy volunteers (mean age: 30 ± 4 years; 13 men) underwent 0.55 T abdominal magnetic resonance imaging from 07/2024-09/2024. Conventional and DL-enhanced DWI were acquired with two b values (b50 s/mm<sup>2</sup>, b800 s/mm<sup>2</sup>). Three fellowship-trained abdominal radiologists rated the scans for quality parameters and artifacts (Likert scales, 1-5). Quantitative b50 and b800 signal intensity and ADC values were measured in the right and left liver lobes, spleen and Th12 vertebral body by one reader. Interreader reliability, intermethod concordance, and Bland-Altman analyses were performed.</div></div><div><h3>Results</h3><div>DL-enhanced scans demonstrated higher quality across all diffusion weightings (all: <em>P</em> < .001): b50 images (median: 4 [IQR: 4-5] vs 3 [3-4]), b800 images (4 [4-4] vs 3 [3-3]), and ADC maps (4 [4-4] vs 3 [3-3]) with good or better interreader agreement (κ ≥ 0.63). Image noise, spatial resolution, organ sharpness and artifacts were rated better for DL-enhanced DWI (all: <em>P</em> < .001).</div><div>Mean b50 and b800 signal intensities were lower with DL reconstruction (all: <em>P</em> < .001). ADC values showed at least strong intermethod correlation (r ≥ 0.61; all: <em>P</em> < .001) with mean differences of 0.3%-4.5% in normal tissues. Bland-Altman plots and interchangeability analysis confirmed intermethod ADC value deviations within expected margins of ±20% (all: <em>P</em> < .001).</div></div><div><h3>Conclusion</h3><div>DL superresolution reconstruction enables faster and higher quality abdominal DWI at 0.55 T without relevant ADC deviations, supporting interchangeability with conventional DWI.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 815-823"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145758333","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2026-03-02DOI: 10.1016/j.acra.2026.01.034
Mohiuddin Hadi MD , Joshua Brown MD, PhD , Florence X. Doo MD, MA , Christopher Straus MD , Anna Rozenshtein MD, MPH , Michele Retrouvey MD
Radiology is increasingly defined by its reliance on complex data, advanced computation, and secure information exchange. Among the emerging technologies are quantum computing and blockchain, which, though originating outside the traditional radiologic workflow, have the potential to transform how imaging data is processed, analyzed, and shared. Quantum computing could accelerate key tasks such as image reconstruction, radiomics analysis, and AI model training by leveraging superposition and entanglement. Blockchain offers a decentralized architecture for ensuring data integrity, provenance, and verifiable access in contexts such as multi-reader interpretations, patient-controlled imaging exchange, and research tracking. These technologies may also converge, raising the importance of developing quantum-resistant blockchain solutions to safeguard privacy and trust as quantum systems mature. As part 5 of the Radiology Research Alliance (RRA) review series on emerging technologies, this paper provides an accessible overview of quantum computing and blockchain, equipping radiologists with the tools to critically engage with the digital innovations shaping the future of the specialty.
{"title":"An RRA Perspective on Quantum Computing and Blockchain in Radiology: Emerging Paradigms for Data Integrity and Advanced Computation","authors":"Mohiuddin Hadi MD , Joshua Brown MD, PhD , Florence X. Doo MD, MA , Christopher Straus MD , Anna Rozenshtein MD, MPH , Michele Retrouvey MD","doi":"10.1016/j.acra.2026.01.034","DOIUrl":"10.1016/j.acra.2026.01.034","url":null,"abstract":"<div><div>Radiology is increasingly defined by its reliance on complex data, advanced computation, and secure information exchange. Among the emerging technologies are quantum computing and blockchain, which, though originating outside the traditional radiologic workflow, have the potential to transform how imaging data is processed, analyzed, and shared. Quantum computing could accelerate key tasks such as image reconstruction, radiomics analysis, and AI model training by leveraging superposition and entanglement. Blockchain offers a decentralized architecture for ensuring data integrity, provenance, and verifiable access in contexts such as multi-reader interpretations, patient-controlled imaging exchange, and research tracking. These technologies may also converge, raising the importance of developing quantum-resistant blockchain solutions to safeguard privacy and trust as quantum systems mature. As part 5 of the Radiology Research Alliance (RRA) review series on emerging technologies, this paper provides an accessible overview of quantum computing and blockchain, equipping radiologists with the tools to critically engage with the digital innovations shaping the future of the specialty.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 653-661"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147357012","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2025-11-12DOI: 10.1016/j.acra.2025.10.013
Zhong Yang , Wangyang Sun , Yanran Jiang , Kejia Guo , Tingting Lin , Yulan Chen , Chuanbin Wang , Bin Shi , Mengshi Fang , Chao Wei MD
<div><h3>Rationale and Objectives</h3><div>This study aims to develop interpretable machine learning (ML) models by integrating conventional magnetic resonance imaging (MRI) features and radiomics to preoperatively differentiate uterine sarcoma (US) and atypical leiomyoma (ALM).</div></div><div><h3>Materials and Methods</h3><div>In this retrospective study, 160 patients (47 US, 113 ALM) were randomized into training (<em>n<!--> </em>=<!--> <!-->112) and test (<em>n<!--> </em>=<!--> <!-->48) cohorts. Two blinded radiologists assessed 10 MRI features from pelvic MRI examinations, including tumor border morphology, T2-weighted image (T2WI) signal heterogeneity, uterine endometrial cavity, apparent diffusion coefficient (ADC) value, and other features. Significant MRI features were identified through univariable and multivariable logistic regression analyses. Radiomics features were extracted from axial T2WI and diffusion-weighted imaging (DWI) sequences, with least absolute shrinkage and selection operator regression identifying four discriminative features for radiomic score (radscore) calculation. Five ML models are as follows: logistic regression (LR), random forest (RF), eXtreme gradient boosting (XGBoost), support vector machine (SVM), and Gaussian Naive Bayes (GNB) were trained using significant MRI predictors and radscore. Model performance was evaluated via area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The SHapley Additive exPlanation (SHAP) framework provided interpretable visualizations of feature contributions.</div></div><div><h3>Results</h3><div>Multivariable analysis identified four MRI discriminators as follows: heterogeneous hyperintensity on T2WI (odds ratio [OR]<!--> <!-->=<!--> <!-->43.767, <em>P<!--> </em>=<!--> <!-->0.021), ill-defined tumor border (OR<!--> <!-->=<!--> <!-->4.887, <em>P<!--> </em>=<!--> <!-->0.038), interrupted uterine cavity (OR<!--> <!-->=<!--> <!-->15.947, <em>P<!--> </em>=<!--> <!-->0.003), and low ADC values (OR<!--> <!-->=<!--> <!-->0.026, <em>P<!--> </em>=<!--> <!-->0.009). The XGBoost model achieved superior performance, with AUCs of 0.991 (95% confidence interval [CI]: 0.978–1.000) and 0.909 (95% CI: 0.822–0.995) in training and test cohorts, respectively. SHAP analysis highlighted ADC value as the most influential predictor, followed by tumor border, signal intensity on T2WI, radscore, and uterine endometrial cavity. DCA confirmed clinical utility across probability thresholds, and calibration curves demonstrated strong agreement between predicted and observed outcomes.</div></div><div><h3>Conclusion</h3><div>Interpretable ML models integrating MRI biomarkers and radiomics provide a transparent and clinically actionable tool for preoperative differentiation of US and ALM. By quantifying feature contributions through SHAP and providing a transparent SHAP value, this framework bridges the "black-box" gap in ML, fostering clinicians tr
{"title":"Interpretable Machine Learning Model for Differentiating Uterine Sarcoma From Atypical Leiomyoma Based on Conventional MRI Features and Radiomics","authors":"Zhong Yang , Wangyang Sun , Yanran Jiang , Kejia Guo , Tingting Lin , Yulan Chen , Chuanbin Wang , Bin Shi , Mengshi Fang , Chao Wei MD","doi":"10.1016/j.acra.2025.10.013","DOIUrl":"10.1016/j.acra.2025.10.013","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>This study aims to develop interpretable machine learning (ML) models by integrating conventional magnetic resonance imaging (MRI) features and radiomics to preoperatively differentiate uterine sarcoma (US) and atypical leiomyoma (ALM).</div></div><div><h3>Materials and Methods</h3><div>In this retrospective study, 160 patients (47 US, 113 ALM) were randomized into training (<em>n<!--> </em>=<!--> <!-->112) and test (<em>n<!--> </em>=<!--> <!-->48) cohorts. Two blinded radiologists assessed 10 MRI features from pelvic MRI examinations, including tumor border morphology, T2-weighted image (T2WI) signal heterogeneity, uterine endometrial cavity, apparent diffusion coefficient (ADC) value, and other features. Significant MRI features were identified through univariable and multivariable logistic regression analyses. Radiomics features were extracted from axial T2WI and diffusion-weighted imaging (DWI) sequences, with least absolute shrinkage and selection operator regression identifying four discriminative features for radiomic score (radscore) calculation. Five ML models are as follows: logistic regression (LR), random forest (RF), eXtreme gradient boosting (XGBoost), support vector machine (SVM), and Gaussian Naive Bayes (GNB) were trained using significant MRI predictors and radscore. Model performance was evaluated via area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA). The SHapley Additive exPlanation (SHAP) framework provided interpretable visualizations of feature contributions.</div></div><div><h3>Results</h3><div>Multivariable analysis identified four MRI discriminators as follows: heterogeneous hyperintensity on T2WI (odds ratio [OR]<!--> <!-->=<!--> <!-->43.767, <em>P<!--> </em>=<!--> <!-->0.021), ill-defined tumor border (OR<!--> <!-->=<!--> <!-->4.887, <em>P<!--> </em>=<!--> <!-->0.038), interrupted uterine cavity (OR<!--> <!-->=<!--> <!-->15.947, <em>P<!--> </em>=<!--> <!-->0.003), and low ADC values (OR<!--> <!-->=<!--> <!-->0.026, <em>P<!--> </em>=<!--> <!-->0.009). The XGBoost model achieved superior performance, with AUCs of 0.991 (95% confidence interval [CI]: 0.978–1.000) and 0.909 (95% CI: 0.822–0.995) in training and test cohorts, respectively. SHAP analysis highlighted ADC value as the most influential predictor, followed by tumor border, signal intensity on T2WI, radscore, and uterine endometrial cavity. DCA confirmed clinical utility across probability thresholds, and calibration curves demonstrated strong agreement between predicted and observed outcomes.</div></div><div><h3>Conclusion</h3><div>Interpretable ML models integrating MRI biomarkers and radiomics provide a transparent and clinically actionable tool for preoperative differentiation of US and ALM. By quantifying feature contributions through SHAP and providing a transparent SHAP value, this framework bridges the \"black-box\" gap in ML, fostering clinicians tr","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 952-962"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145514797","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Detecting Early Ischemic Changes (EIC) on noncontrast computed tomography (NCCT) is essential in patient selection for reperfusion therapy in acute ischemic stroke (AIS). However, identifying these subtle changes remains challenging due to their variable presentation, dependence on reader expertise, and significant interobserver variability. Therefore, an objective method for identifying and quantifying early ischemic brain damage is needed to assist clinicians, particularly in resource-limited settings.
Recent advances in artificial intelligence (AI), including machine learning (ML) and deep learning (DL), have opened new opportunities to enhance stroke diagnosis by enabling fast, consistent, and accurate analysis of NCCT images.
This review summarizes current AI applications in detecting EIC on NCCT images, focusing on two major developments: (1) the automatic calculation of the Alberta Stroke Program Early Computed Tomography Score (ASPECTS), which facilitates automated tracing of regions of interest (ROIs) and quantification of hypoattenuation to objectively assess ischemic damage, and (2) DL-based EIC detection approaches, supported by large-scale datasets.
We highlight the potential of these innovations to complement clinical expertise, streamline workflows, and improve patient outcomes. We discuss the methodologies, performance metrics, and limitations of existing AI models. By synthesizing the latest research, this paper explores AI’s transformative role in AIS management and outlines future directions for innovation in this rapidly evolving field.
{"title":"Advancing Stroke Diagnosis: A Comprehensive Review of Artificial Intelligence in Detecting Early Ischemic Changes on Noncontrast CT (NCCT)","authors":"Ines Ben Alaya , Fethi Felhi , Mariem Messelmani , Salam Labidi","doi":"10.1016/j.acra.2025.11.015","DOIUrl":"10.1016/j.acra.2025.11.015","url":null,"abstract":"<div><div>Detecting Early Ischemic Changes (EIC) on noncontrast computed tomography (NCCT) is essential in patient selection for reperfusion therapy in acute ischemic stroke (AIS). However, identifying these subtle changes remains challenging due to their variable presentation, dependence on reader expertise, and significant interobserver variability. Therefore, an objective method for identifying and quantifying early ischemic brain damage is needed to assist clinicians, particularly in resource-limited settings.</div><div>Recent advances in artificial intelligence (AI), including machine learning (ML) and deep learning (DL), have opened new opportunities to enhance stroke diagnosis by enabling fast, consistent, and accurate analysis of NCCT images.</div><div>This review summarizes current AI applications in detecting EIC on NCCT images, focusing on two major developments: (1) the automatic calculation of the Alberta Stroke Program Early Computed Tomography Score (ASPECTS), which facilitates automated tracing of regions of interest (ROIs) and quantification of hypoattenuation to objectively assess ischemic damage, and (2) DL-based EIC detection approaches, supported by large-scale datasets.</div><div>We highlight the potential of these innovations to complement clinical expertise, streamline workflows, and improve patient outcomes. We discuss the methodologies, performance metrics, and limitations of existing AI models. By synthesizing the latest research, this paper explores AI’s transformative role in AIS management and outlines future directions for innovation in this rapidly evolving field.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 1060-1069"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145607129","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Spondylolisthesis, a vertebral displacement condition affecting 5–26% of adults, poses a significant health risk to the population. Artificial Intelligence (AI), has emerged as a tool for enhancing diagnostic accuracy. However, the heterogeneity in model performances requires a synthesis of existing evidence.
Materials and Methods
This study evaluated the diagnostic accuracy of AI models for the detection of spondylolisthesis across multiple imaging modalities. Following PRISMA PubMed, Scopus, Embase, Web of Science, including 24 studies (21 for meta-analysis) with 8029 observations. Inclusion criteria focused on original studies using standalone deep learning (DL) models with reported diagnostic metrics. Quality assessment was performed using Quality Assessment of Diagnostic Accuracy Studies-2, and statistical analysis employed random-effects meta-analysis.
Results
AI models demonstrated high diagnostic performance, with a pooled sensitivity of 94.7% (95% CI: 92.6–96.2%) and specificity of 97.1% (95% CI: 95.0–98.4%). The area under the curve (AUC) was 0.979, indicating robust discriminative ability. MRI-based models slightly outperformed radiography models (sensitivity: 95.71% vs. 94.95%; specificity: 98.38% vs. 96.80%), though differences were nonsignificant (p = 0.651). Classification models significantly surpassed detection-focused models (p = 0.026), while biomechanical feature-based models and DL image processing models showed comparable performance (p = 0.264). Notably, models like FAR networks and YOLOv8 achieved high accuracy (89–98%) in grading and localization tasks.
Conclusions
AI models show considerable diagnostic accuracy for spondylolisthesis, underscoring their potential as clinical adjunctive tools. However, considerable heterogeneity highlights the need for standardized studies. These findings advocate for integrating AI into diagnostic workflows, particularly in resource-limited settings, while urging further research to ensure real-world applicability.
理由和目的：椎体滑脱是一种影响5-26%成年人的椎体移位疾病，对人群构成重大健康风险。人工智能（AI）已经成为提高诊断准确性的工具。然而，模型性能的异质性需要综合现有证据。材料和方法：本研究评估了人工智能模型在多种成像方式下检测脊柱滑脱的诊断准确性。追踪PRISMA PubMed， Scopus, Embase, Web of Science，包括24项研究（21项为荟萃分析），8029项观察结果。纳入标准侧重于使用独立深度学习（DL）模型和报告的诊断指标的原始研究。质量评估采用诊断准确性研究质量评估-2，统计分析采用随机效应荟萃分析。结果：人工智能模型表现出较高的诊断性能，合并敏感性为94.7% (95% CI: 92.6-96.2%)，特异性为97.1% （95% CI: 95.0-98.4%）。曲线下面积（AUC）为0.979，判别能力较强。基于mri的模型略优于x线摄影模型（灵敏度：95.71% vs. 94.95%；特异性：98.38% vs. 96.80%），但差异无统计学意义（p = 0.651）。分类模型显著优于以检测为中心的模型（p = 0.026），而基于生物力学特征的模型和深度学习图像处理模型的性能相当（p = 0.264）。值得注意的是，像FAR网络和YOLOv8这样的模型在分级和定位任务中实现了很高的准确率（89-98%）。结论：人工智能模型对脊柱滑脱的诊断具有相当高的准确性，强调了其作为临床辅助工具的潜力。然而，相当大的异质性突出了标准化研究的必要性。这些发现提倡将人工智能集成到诊断工作流程中，特别是在资源有限的情况下，同时敦促进一步研究以确保现实世界的适用性。
{"title":"Evaluating the Diagnostic Accuracy of Artificial Intelligence in Spondylolisthesis Detection: A Systematic Review and Meta-analysis","authors":"Mohammad-Taha Pahlevan-Fallahy MD, MPH , Amir-Mohammad Asgari MD , Alireza Soltani Khaboushan MD , Majid Chalian , Farhad Shaker MD , Parnian Yari MD , Sara Haseli","doi":"10.1016/j.acra.2025.11.002","DOIUrl":"10.1016/j.acra.2025.11.002","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>Spondylolisthesis, a vertebral displacement condition affecting 5–26% of adults, poses a significant health risk to the population. Artificial Intelligence (AI), has emerged as a tool for enhancing diagnostic accuracy. However, the heterogeneity in model performances requires a synthesis of existing evidence.</div></div><div><h3>Materials and Methods</h3><div>This study evaluated the diagnostic accuracy of AI models for the detection of spondylolisthesis across multiple imaging modalities. Following PRISMA PubMed, Scopus, Embase, Web of Science, including 24 studies (21 for meta-analysis) with 8029 observations. Inclusion criteria focused on original studies using standalone deep learning (DL) models with reported diagnostic metrics. Quality assessment was performed using Quality Assessment of Diagnostic Accuracy Studies-2, and statistical analysis employed random-effects meta-analysis.</div></div><div><h3>Results</h3><div>AI models demonstrated high diagnostic performance, with a pooled sensitivity of 94.7% (95% CI: 92.6–96.2%) and specificity of 97.1% (95% CI: 95.0–98.4%). The area under the curve (AUC) was 0.979, indicating robust discriminative ability. MRI-based models slightly outperformed radiography models (sensitivity: 95.71% vs. 94.95%; specificity: 98.38% vs. 96.80%), though differences were nonsignificant (<em>p</em> = 0.651). Classification models significantly surpassed detection-focused models (<em>p</em> = 0.026), while biomechanical feature-based models and DL image processing models showed comparable performance (<em>p</em> = 0.264). Notably, models like FAR networks and YOLOv8 achieved high accuracy (89–98%) in grading and localization tasks.</div></div><div><h3>Conclusions</h3><div>AI models show considerable diagnostic accuracy for spondylolisthesis, underscoring their potential as clinical adjunctive tools. However, considerable heterogeneity highlights the need for standardized studies. These findings advocate for integrating AI into diagnostic workflows, particularly in resource-limited settings, while urging further research to ensure real-world applicability.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 1034-1048"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145607077","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2025-11-29DOI: 10.1016/j.acra.2025.11.022
Sean Raj MD, MBA, Chief Medical Officer and Chief Innovation Officer, Barry Sadegi MD, John Simon MD
Rationale and Objectives
This study investigates the efficacy of a deep learning-based artificial intelligence (AI) model in detecting pediatric fractures on musculoskeletal (MSK) radiographs and assesses the impact of AI-assistance on the performance of radiologists.
Materials and Methods
In Phase 1, the performance of the AI model was evaluated on 3016 MSK pediatric radiographs from 4 imaging centers in the US. Ground truth was established by consensus of pediatric radiologists. Phase 2 was a retrospective multi-reader, multi-center (MRMC) study using 189 cases. Twenty readers participated in two separate reading sessions evaluating for fracture, with and without AI assistance, with a one-month washout period.
Results
The AI model achieved a high standalone performance with accuracy (0.94), sensitivity (0.96), and specificity (0.86). Subgroup analysis revealed that the model maintained high performance across study types and confounders, including age (Se>0.94), gender (Se>0.96), anatomical region (Se>0.93), and fracture types (Se>0.93). With AI assistance, reader accuracy increased significantly from 0.93 to 0.96 (p < 0.05), sensitivity significantly improved from 0.86 to 0.93 (p < 0.05), and specificity improved from 0.94 to 0.95. The average reading time per exam was shortened by 26.1% with AI assistance.
Conclusion
The AI model's high accuracy in detecting pediatric fractures underscores its significant clinical utility. The integration of this tool enhanced overall radiologist performance and boosted the diagnostic confidence among non-specialist readers.
{"title":"Enhancing Pediatric Fracture Detection: Multicenter Evaluation of a Deep Learning AI Model and Its Impact on Radiologist Performance","authors":"Sean Raj MD, MBA, Chief Medical Officer and Chief Innovation Officer, Barry Sadegi MD, John Simon MD","doi":"10.1016/j.acra.2025.11.022","DOIUrl":"10.1016/j.acra.2025.11.022","url":null,"abstract":"<div><h3>Rationale and Objectives</h3><div>This study investigates the efficacy of a deep learning-based artificial intelligence (AI) model in detecting pediatric fractures on musculoskeletal (MSK) radiographs and assesses the impact of AI-assistance on the performance of radiologists.</div></div><div><h3>Materials and Methods</h3><div>In Phase 1, the performance of the AI model was evaluated on 3016 MSK pediatric radiographs from 4 imaging centers in the US. Ground truth was established by consensus of pediatric radiologists. Phase 2 was a retrospective multi-reader, multi-center (MRMC) study using 189 cases. Twenty readers participated in two separate reading sessions evaluating for fracture, with and without AI assistance, with a one-month washout period.</div></div><div><h3>Results</h3><div>The AI model achieved a high standalone performance with accuracy (0.94), sensitivity (0.96), and specificity (0.86). Subgroup analysis revealed that the model maintained high performance across study types and confounders, including age (Se>0.94), gender (Se>0.96), anatomical region (Se>0.93), and fracture types (Se>0.93). With AI assistance, reader accuracy increased significantly from 0.93 to 0.96 (p < 0.05), sensitivity significantly improved from 0.86 to 0.93 (p < 0.05), and specificity improved from 0.94 to 0.95. The average reading time per exam was shortened by 26.1% with AI assistance.</div></div><div><h3>Conclusion</h3><div>The AI model's high accuracy in detecting pediatric fractures underscores its significant clinical utility. The integration of this tool enhanced overall radiologist performance and boosted the diagnostic confidence among non-specialist readers.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 1121-1129"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145649731","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2025-12-01DOI: 10.1016/j.acra.2025.11.024
Lin Mu , Yao Liu , Yunming Xie , Haoyu Liu , Ke Liu , Zheng Miao , Han Xue , Mingyang Li , Dong Dong , Huimao Zhang
Rationale and objectives
To develop and evaluate a femoral neck fracture (FNF) pipeline model for diagnosing fracture stability and aiding surgical decision-making.
Materials and Methods
Patients with confirmed FNFs were enrolled in the study. An automatic segmentation algorithm was employed to initially delineate fracture-displaced regions revealed using CT images, with subsequent manual refinement. A logistic-regression model was first trained on selected radiomic features to generate a Rad-score for fracture-stability classification. The Rad_score was then fed into a downstream model to guide surgical decision-making. The internal and external validation with multi-center data were used to assess the generalizability of the pipeline model.
Results
The internal dataset for fracture stability and surgical decision-making included 624 and 410 patients, respectively. The corresponding external test sets, included 364 and 186 patients enrolled from 32 centers. The radiomics model for FNF stability exhibited robust performance, achieving an area under the curve (AUC) of 0.905 (95% confidence interval [CI]: 0.853–0.944) and 0.821 (95% CI: 0.778–0.859) for the internal and external test sets, respectively. The AUCs for the surgical decision-making models were 0.881 (95% CI: 0.810–0.932) and 0.820 (95% CI: 0.757–0.873) for the internal and external test sets, respectively.
Conclusion
The radiomics pipeline model exhibited robust performance in classifying fracture stability and aiding surgical decision-making in the test sets across 33 centers. Our model incorporates explainable artificial intelligence in fracture quantification analysis, supporting doctors in making objective clinical decisions.
{"title":"From Imaging to Intervention: A Multicenter-Validated Radiomics Pipeline for Guiding Femoral Neck Fracture Surgical Management","authors":"Lin Mu , Yao Liu , Yunming Xie , Haoyu Liu , Ke Liu , Zheng Miao , Han Xue , Mingyang Li , Dong Dong , Huimao Zhang","doi":"10.1016/j.acra.2025.11.024","DOIUrl":"10.1016/j.acra.2025.11.024","url":null,"abstract":"<div><h3>Rationale and objectives</h3><div>To develop and evaluate a femoral neck fracture (FNF) pipeline model for diagnosing fracture stability and aiding surgical decision-making.</div></div><div><h3>Materials and Methods</h3><div>Patients with confirmed FNFs were enrolled in the study. An automatic segmentation algorithm was employed to initially delineate fracture-displaced regions revealed using CT images, with subsequent manual refinement. A logistic-regression model was first trained on selected radiomic features to generate a Rad-score for fracture-stability classification. The Rad_score was then fed into a downstream model to guide surgical decision-making. The internal and external validation with multi-center data were used to assess the generalizability of the pipeline model.</div></div><div><h3>Results</h3><div>The internal dataset for fracture stability and surgical decision-making included 624 and 410 patients, respectively. The corresponding external test sets, included 364 and 186 patients enrolled from 32 centers. The radiomics model for FNF stability exhibited robust performance, achieving an area under the curve (AUC) of 0.905 (95% confidence interval [CI]: 0.853–0.944) and 0.821 (95% CI: 0.778–0.859) for the internal and external test sets, respectively. The AUCs for the surgical decision-making models were 0.881 (95% CI: 0.810–0.932) and 0.820 (95% CI: 0.757–0.873) for the internal and external test sets, respectively.</div></div><div><h3>Conclusion</h3><div>The radiomics pipeline model exhibited robust performance in classifying fracture stability and aiding surgical decision-making in the test sets across 33 centers. Our model incorporates explainable artificial intelligence in fracture quantification analysis, supporting doctors in making objective clinical decisions.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 1049-1059"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145662641","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-03-01Epub Date: 2026-01-16DOI: 10.1016/j.acra.2025.12.050
Cunke Miao , Houzhang Sun , Fei Yao , Tianle Hong , Zedong Ren , Yuandi Zhuang , Qi Lin , Shuying Bian , Yunjun Yang , Yezhi Lin
Background
Prostate cancer is the second most common cancer in men, with rising mortality rates necessitating precise risk stratification. High-invasive biological features—specifically International Society of Urological Pathology (ISUP) grade, extracapsular extension (EPE), and positive surgical margins (PSM)—are critical for guiding treatment but are difficult to detect due to tumor heterogeneity. Current imaging modalities, including 18F-PSMA-1007 PET/CT and multiparametric MRI (mpMRI), have limitations in fully capturing these features. This study aims to develop a few-shot deep learning model (CL-MGNET) that integrates multimodal imaging and clinical data to predict high-risk biological features, optimizing performance even with limited training data.
Materials and Methods
This retrospective, multicenter study analyzed data from 377 patients: 341 from a primary medical center (Center A) and 36 from an independent external validation cohort (Center B). The study utilized multimodal inputs (PET/CT, mpMRI) and clinical variables to predict ISUP grade, EPE, and PSM. A specialized few-shot deep learning network, CL-MGNET, was designed to fuse these data sources. The model was trained using a restricted subset of 30 patients and subsequently evaluated on both internal and external test sets to assess generalizability across different centers.
Results
CL-MGNET demonstrated excellent performance in predicting high-invasive biological features (defined as the presence of at least one high-risk feature: ISUP ≥ 3, EPE, or PSM), achieving an internal test AUC of 0.877 and an external validation AUC of 0.872, which significantly outperformed the clinical model with an AUC of 0.792. The model surpassed both single-modality models (PET/CT, mpMRI) and the clinical model. Furthermore, CL-MGNET exhibited strong generalization capability, effectively predicting various high-risk biological features. When clinical variables were integrated, the model's performance improved significantly, exceeding traditional methods.
Conclusion
The CL-MGNET model, leveraging multimodal imaging data and clinical variables with a few-shot learning approach, successfully predicts high-invasive biological features of prostate cancer with high accuracy, even with limited data. The model's performance across different biological features and medical centers shows its robust generalizability. This method holds great promise for improving prostate cancer diagnosis and risk prediction in data-limited environments.
{"title":"Generalizable Deep Learning for Prostate Cancer Risk Stratification: Multicenter Study Integrating 18F-PSMA-1007 PET/CT and mpMRI","authors":"Cunke Miao , Houzhang Sun , Fei Yao , Tianle Hong , Zedong Ren , Yuandi Zhuang , Qi Lin , Shuying Bian , Yunjun Yang , Yezhi Lin","doi":"10.1016/j.acra.2025.12.050","DOIUrl":"10.1016/j.acra.2025.12.050","url":null,"abstract":"<div><h3>Background</h3><div>Prostate cancer is the second most common cancer in men, with rising mortality rates necessitating precise risk stratification. High-invasive biological features—specifically International Society of Urological Pathology (ISUP) grade, extracapsular extension (EPE), and positive surgical margins (PSM)—are critical for guiding treatment but are difficult to detect due to tumor heterogeneity. Current imaging modalities, including 18F-PSMA-1007 PET/CT and multiparametric MRI (mpMRI), have limitations in fully capturing these features. This study aims to develop a few-shot deep learning model (CL-MGNET) that integrates multimodal imaging and clinical data to predict high-risk biological features, optimizing performance even with limited training data.</div></div><div><h3>Materials and Methods</h3><div>This retrospective, multicenter study analyzed data from 377 patients: 341 from a primary medical center (Center A) and 36 from an independent external validation cohort (Center B). The study utilized multimodal inputs (PET/CT, mpMRI) and clinical variables to predict ISUP grade, EPE, and PSM. A specialized few-shot deep learning network, CL-MGNET, was designed to fuse these data sources. The model was trained using a restricted subset of 30 patients and subsequently evaluated on both internal and external test sets to assess generalizability across different centers.</div></div><div><h3>Results</h3><div>CL-MGNET demonstrated excellent performance in predicting high-invasive biological features (defined as the presence of at least one high-risk feature: ISUP ≥ 3, EPE, or PSM), achieving an internal test AUC of 0.877 and an external validation AUC of 0.872, which significantly outperformed the clinical model with an AUC of 0.792. The model surpassed both single-modality models (PET/CT, mpMRI) and the clinical model. Furthermore, CL-MGNET exhibited strong generalization capability, effectively predicting various high-risk biological features. When clinical variables were integrated, the model's performance improved significantly, exceeding traditional methods.</div></div><div><h3>Conclusion</h3><div>The CL-MGNET model, leveraging multimodal imaging data and clinical variables with a few-shot learning approach, successfully predicts high-invasive biological features of prostate cancer with high accuracy, even with limited data. The model's performance across different biological features and medical centers shows its robust generalizability. This method holds great promise for improving prostate cancer diagnosis and risk prediction in data-limited environments.</div></div>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":"33 3","pages":"Pages 1107-1120"},"PeriodicalIF":3.9,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145994724","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号