Academic Radiology最新文献

{"title":"Added Value of CT Bone Reading Software in the Detection of Subtle Bone Metastases From Breast Cancer: Comparison with Standard Thorax CT.","authors":"Derya Karabulut, Aykut Alkan","doi":"10.1016/j.acra.2026.09.046","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.046","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>The aim of this study is to evaluate the additional diagnostic value of CT Bone Reading software in detecting rib metastases on thoracic computed tomography in patients with breast cancer and to examine its effectiveness on lesion detection, evaluation time, and inter-reader agreement.</p><p><strong>Materials and methods: </strong>This retrospective study included 110 patients with breast cancer and bone metastases who underwent both thoracic computed tomography (CT) and positron emission tomography-CT scans between January 2022 and January 2026. For lesion-level assessment, suspected rib lesions were also evaluated using available longitudinal follow-up imaging. Two radiologists independently evaluated conventional CT and CT Bone Reading images in separate reading sessions at least 2 weeks apart. Inter-reader agreement was assessed using the intraclass correlation coefficient (ICC) and Cohen's κ statistic.</p><p><strong>Results: </strong>In lesion-based analysis with conventional CT, Reader 1 detected 510 lesions (81.6%) and Reader 2 detected 577 lesions (92.3%), while CT Bone Reading allowed for the detection of an additional 115 (22.5%) and 48 (8.3%) lesions for Readers 1 and 2, respectively. The additionally detected lesions demonstrated a non-uniform distribution, with a marked predominance in the mid-rib region (levels 5th-8th). CT Bone Reading-assisted evaluation enabled the detection of significantly smaller lesions compared to routine CT for both readers (Reader 1: 8.0 mm vs 5.0 mm; Reader 2: 7.0 mm vs 5.0 mm; P < 0.001 for both). Evaluation time was significantly decreased from 129.22 ± 27.22 s to 64.94 ± 17.16 s for Reader 1 and from 122.99 ± 26.04 s to 64.44 ± 18.71 s for Reader 2 (P < 0.001 for both). Inter-reader agreement was excellent for measurements obtained by CT Bone Reading-derived evaluation (ICC ≥0.965; κ ≥0.947).</p><p><strong>Conclusion: </strong>CT Bone Reading software increases the detection rates of rib lesions, reduces interpretation times, and demonstrates excellent inter-reader reproducibility compared to conventional CT assessment, with less experienced readers achieving the highest incremental throughput. By providing a comprehensive and continuous view of the ribs, this technique has the potential to improve diagnostic performance without additional radiation exposure, particularly in anatomically complex regions such as the mid ribs. Ultimately, this approach can enhance workflow efficiency during routine clinical practices.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149169557","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}

{"title":"From Response Assessment to Biological Insight: PET/CT and Artificial Intelligence in Metastatic Breast Cancer.","authors":"Yi Liang, Kun Wang","doi":"10.1016/j.acra.2026.09.045","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.045","url":null,"abstract":"","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149169317","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}

{"title":"Utility of OpenEvidence and ChatGPT in Selecting Neuroimaging Based on Appropriateness Criteria.","authors":"Siddhant Kumarapuram, Medha Satti, Shirleen Xu, Jay Kavia, Ashwin Bhaskar, Aafreen Azmi, Alberto F Goldszal, Andrew Soliman, Srihari Sundararajan, Gaurav Gupta, Sudipta Roychowdhury","doi":"10.1016/j.acra.2026.09.049","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.049","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>Large Language models (LLMs) are novel tools with promising applications in imaging selection for clinical scenarios, but information is limited on the accuracy of an LLM trained on a medically specific dataset compared to a generalized dataset. In this analysis, we assessed the accuracy of the imaging recommendations by LLMs ChatGPT-3.5, ChatGPT-5.0, and OpenEvidence for neuroradiology, using the American College of Radiology (ACR) Appropriateness Criteria as the reference standard.</p><p><strong>Materials and methods: </strong>Prompts from the ACR Appropriateness Criteria priority clinical areas within neurology were used to build open-ended (OE) or select-all-that-apply (SATA) questions. Each question was posed three times, and the responses were scored against ACR guidelines. The performance of each LLM on OE and SATA was compared, and then was further analyzed by acuity. Mean accuracy scores were calculated by averaging performance across OE and SATA questions for each model. Analysis of variance was then conducted to assess differences in performance among the models.</p><p><strong>Results: </strong>A total of 78 prompts were assessed in 11 priority clinical areas. These areas included cerebrovascular conditions, neurological pain, and spine pathologies. Amongst SATA, Open Evidence performed the best (79.5%) compared to GPT5 (71.2%) and GPT3.5 (44.4%). For the OE prompts, Open Evidence was more accurate (86.3%) than GPT3.5 (69.2%) but not GPT5 (80.6%). Amongst the models, no differences were found based on acuity.</p><p><strong>Conclusion: </strong>Overall, OpenEvidence performed the best for both OE and SATA prompts. ChatGPT-5 performed better than ChatGPT3.5 for SATA prompts. LLM performance was unaffected by the acuity of the case.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149171974","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}

{"title":"Hyperintense Vessel Sign on High-Resolution Vessel Wall MRI and Perfusion Decompensation in Intracranial Atherosclerotic Disease: Burden Versus Proximal Extent.","authors":"Feng Ouyang, Qin Wu, Jie Liu, Shasha Lai, Lan Luo, Wenchao Xiong, Yixin Liu, Lu Sun, Laisheng Pan, Fuqing Zhou, Xianjun Zeng","doi":"10.1016/j.acra.2026.09.052","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.052","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>To investigate the relationship between hyperintense vessel sign (HVS) on high-resolution magnetic resonance vessel wall imaging (HR-VWI) and cerebral perfusion decompensation in patients with intracranial atherosclerotic stenosis or occlusion.</p><p><strong>Materials and methods: </strong>This retrospective cohort included 128 patients with unilateral intracranial atherosclerotic stenosis or occlusion undergoing HR-VWI and computed tomography perfusion. HVS was scored by vessel count and classified by extent (type 0, none; type 1, M3-M4; type 2, M1-M2); local collaterals were assessed on time-of-flight magnetic resonance angiography. Prespecified Firth-penalized regression, bootstrap-validated threshold modeling, hierarchical likelihood-ratio tests, and collateral-phenotype analysis compared grade 2 versus grade 1.</p><p><strong>Results: </strong>The study cohort comprised 128 patients (92 males, 36 females), with a mean age of 55.08 ± 10.58 years. 85 patients were grade 1, 43 grade 2. HVS score was independently associated with decompensation (odds ratio [OR], 1.09, P = 0.04), with an exploratory cutoff at 8 (bootstrap 95% confidence interval [CI], 6-8; adjusted OR, 6.93, P < 0.01). Grade-2 proportions differed across type 0/1/2 (25.00%/22.67%/63.64%); type 2 exceeded type 1 (adjusted OR 8.06, P < 0.01). Classification, not the score, added incremental value (likelihood ratio P < 0.01 vs reverse P = 0.15). Type 1 with collateral present marked compensation (5.88%; negative predictive value 0.94); type 2 with collateral absent marked decompensation (66.67%; adjusted OR = 33.90, P < 0.01). Score ≥8 gave 0.65 sensitivity and 0.80 specificity; combined-model area under the curve was 0.83 (95% CI, 0.75-0.90; optimism-corrected 0.79).</p><p><strong>Conclusion: </strong>HVS score provides an exploratory screening cutoff, whereas proximal HVS extension into M1-M2 is strongly associated with decompensation; distal-limited HVS with local collateral formation marks compensation. Prospective multicenter validation is required before clinical implementation.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149148324","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}

{"title":"Retrospective Evaluation of the D-Dimer in Patients Undergoing CTPA for PE Evaluation.","authors":"Hannah E Johnson, Christian J Konopka, Kristin Delfino, Teresa Wilson, Scott Long","doi":"10.1016/j.acra.2026.09.014","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.014","url":null,"abstract":"<p><strong>Objective: </strong>Computed tomography pulmonary angiogram (CTPA) is the choice imaging modality for diagnosis of pulmonary embolism (PE). Studies have demonstrated more appropriate use of CTPA following implementation of the Wells or Revised Geneva scores and/or D-dimer level thresholds into clinical frameworks. During the study period, these measures were not required prior to ordering a CTPA at our institution. The aim of this study was to retrospectively evaluate the institutional utilization and ordering patterns preceding CTPA for evaluation of PE, with focus on the laboratory D-dimer test.</p><p><strong>Methods: </strong>This retrospective chart review included patients that underwent CTPA for suspected PE between January 1, 2022 and December 31, 2022. Data included clinical factors, D-dimer levels, CTPA results, and subspecialty of ordering providers. Statistical analyses were performed via SAS v9.4. P values < 0.05 were considered statistically significant.</p><p><strong>Results: </strong>A total of 660 patients were included in this study, with 54 positives for PE (8.2%). The D-dimer was ordered for 332 patients, and elevated ( ≥ 500 ng/mL) in 20 patients with, and 299 without PE. When analyzed using the institutional threshold as a binary (elevated vs non-elevated), D-dimer levels were not associated with PE on CTPA (p = 0.2099). However, when analyzed as a continuous variable using the standard and age-adjusted thresholds, greater values were associated with PE (p < 0.05). Patients with retrospective high-risk Wells or Revised Geneva scores were more likely to proceed directly to CTPA.</p><p><strong>Discussion: </strong>When analyzed as a continuous variable, elevated D-dimer levels were associated with PE. Our findings also suggest that the D-dimer is more widely used in the diagnostic workup in those at low risk of PE, likely in a \"rule out\" manner. Future work should evaluate D-dimer utilization, laboratory levels, and clinical probability scores of those with suspected PE in groups that did and did not undergo CTPA.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149149478","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}

{"title":"From Adaptation to Integration: Redesigning Academic Radiology for a Transformed Profession.","authors":"Diego Avila Lessa Garcia, Marilia Paes Fortes Diniz Ferreira","doi":"10.1016/j.acra.2026.09.047","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.047","url":null,"abstract":"<p><p>Over the past three decades, radiology has transformed more profoundly than perhaps any other medical specialty. Imaging volumes have multiplied, examinations have grown in complexity, and radiologists have moved from isolated interpreters of static images to central participants in multidisciplinary decision-making. Academic organizations responded to this transformation continuously-but largely through local, successful adaptations layered onto existing structures, rather than through periodic redesign of those structures themselves. This Perspective proposes that many of today's most visible challenges-burnout, workforce instability, fragile subspecialty coverage, and the progressive loss of protected academic time-are not independent problems. They are manifestations of a common underlying pattern, which this Perspective names explicitly as organizational accumulation outpacing organizational integration-decades of local adaptation layered onto existing structures without a matching cycle of structural redesign. Work relative value units, still indispensable for measuring reimbursable clinical activity, illustrate the pattern well: they remain valid for what they were built to measure, while increasingly capturing only a portion of what the modern academic radiologist actually does (1). None of the contemporary pressures facing academic radiology created this imbalance. They exposed it. For decades, an adaptive reserve absorbed the growing gap between professional and organizational evolution-but that reserve resided in people, not in institutions, and it is not unlimited. This Perspective does not propose a specific organizational remedy. It argues that naming this pattern is the necessary first step toward redesigning academic work for the profession radiology has already become. The scarcest resource in academic radiology today may no longer be technology, volume, or expertise. It may be protected intellectual time.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149149130","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}

{"title":"Multimodal Carotid Ultrasound-based Artificial Intelligence Model for Predicting Acute Ischemic Stroke Risk in Patients with Type 2 Diabetes Mellitus.","authors":"Yaning Sun, Bojian Feng, Yanming Zhang, Di Ou, Dangin Pan, Saisai Wang, Fengqin Shen, Xinchen Li, Dong Xu","doi":"10.1016/j.acra.2026.09.036","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.036","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>Patients with type 2 diabetes mellitus (T2DM) are at high risk of acute ischemic stroke (AIS), yet accurate risk stratification remains challenging. This study aimed to develop and validate a multimodal model based on carotid ultrasound for improved AIS risk prediction.</p><p><strong>Materials and methods: </strong>In this multicenter retrospective study, a total of 480 patients with T2DM who underwent carotid ultrasound were recruited from two centers, with one center (n=394) used for model development and the other (n = 86) for independent external testing. A dual-scale deep learning framework based on a Swin Transformer was developed to capture both local plaque features and surrounding vascular context. Deep learning features were integrated with radiomic and clinical features to construct a multimodal model. Model performance was evaluated using the area under the curve (AUC), along with calibration and decision curve analysis.</p><p><strong>Results: </strong>The multimodal model achieved AUCs of 0.952 (95% CI: 0.910-0.984) in the internal cohort and 0.939 (95% CI: 0.838-0.996) in the external cohort, showing improved performance compared to single-modality models. In the reader study, AI-assisted interpretation improved diagnostic performance across all readers, with AUC increases ranging from 0.760 to 0.857 and from 0.868 to 0.929 for senior readers, and from 0.450 to 0.700 and from 0.548 to 0.659 for junior readers. Generalized estimating equation analyses showed significant improvements in diagnostic correctness among both senior and junior readers, with P values of 0.00235 and 0.000343, respectively.</p><p><strong>Conclusion: </strong>A multimodal framework integrating carotid ultrasound with clinical, radiomic, and deep learning features enables improved AIS risk stratification in patients with T2DM.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149149531","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}

{"title":"Asymmetric Complementarity: A Subtype-Specific Strategy via Visual Assessment of DCE-MRI and DWI for Evaluating NAT Efficacy in Breast Cancer.","authors":"Xiuyun Ge, Bin Zheng, Zhen Zhang, Jianle Liang, Wenjing Yuan, Jiarui Xie, Guanxun Cheng, Peng Sun, Rong Huang","doi":"10.1016/j.acra.2026.09.026","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.026","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>To evaluate dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and diffusion-weighted imaging (DWI), based on clinical visual assessment for predicting pathological complete response (pCR) and measuring residual tumor size after neoadjuvant therapy (NAT) across breast cancer subtypes.</p><p><strong>Materials and methods: </strong>This retrospective study included 248 patients. Two radiologists, blinded to pathological results, visually assessed DCE-MRI and DWI (including ADC maps) for radiological complete response (rCR) and measured residual tumor size. Diagnostic performance (vs. pCR) and size agreement (intraclass correlation coefficient, ICC) were compared between sequences and across three molecular subtypes: HR+/HER2-, HER2+, and triple-negative breast cancer (TNBC). Bonferroni correction was applied.</p><p><strong>Results: </strong>DCE-MRI showed higher specificity, positive predictive value (PPV), and accuracy, DWI showed higher sensitivity (all P < 0.05). Subtype analysis: HER2+ exhibited a trend toward symmetric complementarity (DWI-high NPV, DCE-MRI-high PPV; sensitivity difference not significant after correction, adjusted P = 0.066); HR+/HER2- showed DCE-MRI dominance (DWI very low PPV/specificity); TNBC showed no significant differences(n = 30, preliminary). Interobserver κ: 0.77 (DCE-MRI), 0.85 (DWI). Interobserver ICC for size:0.96 (both). DCE-MRI demonstrated superior size concordance with pathology (ICC: 0.91 vs. 0.51, P < 0.001).</p><p><strong>Conclusion: </strong>DCE-MRI and DWI exhibit subtype-dependent asymmetric complementarity. For HER2+, both are complementary (DWI high NPV, DCE-MRI high PPV). For HR+/HER2-, DCE-MRI is the primary anchor (DWI limited). For TNBC, findings are preliminary (larger studies needed). DCE-MRI is more reliable for size measurement. This framework provides a molecular subtype-specific visual interpretation guide for clinical practice.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149118248","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}

{"title":"From Idea to Publication: A Step-by-Step Guide to Original Research in Radiology with AI-Integrated Workflow.","authors":"Isabela Hernandez, Daniel Souza, Cinthia Del Toro, Kristin Parker, Adriana Falcon Vazquez, Gary H Danton, Ty Subhawong, Francesco Alessandrino","doi":"10.1016/j.acra.2026.09.042","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.042","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>Over the past decade, according to the National Resident Matching Program Charting Outcomes data, scholarly output of matched United States Medical Doctor radiology applicants has more than doubled, with applicants reporting an average of 12.0 published scholarly works in 2024 compared to 5.7 in 2014. Despite this shift, structured guidance on how to produce peer-reviewed scholarly work remains limited for radiology trainees, and most existing guides predate the widespread availability of artificial intelligence (AI) tools.</p><p><strong>Materials and methods: </strong>This guide provides a five-step framework, covering choosing a topic of interest, methodology development, presentation of results, manuscript writing, and revision and submission, with explicit integration of AI tools at each stage.</p><p><strong>Results: </strong>A framework for responsible AI use and disclosure is provided.</p><p><strong>Conclusion: </strong>Scientific writing is a teachable skill. AI tools, when used with transparency and integrity, lower the threshold for entry for radiology trainees and should be integrated into scientific writing education. This guide provides a reproducible, AI-integrated framework applicable to trainees at any stage of training and seeking to publish peer-reviewed scholarly works.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149129039","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}

{"title":"Development and Validation of an Interpretable Machine Learning Model Based on Gd-EOB-DTPA-enhanced MRI for Evaluating Small HCC (≤2 cm): A Multicenter Cohort Study.","authors":"Longshan Yang, Xiaojun Zeng, Zhelin Lv, Pengbo Wu, Haitao Wen, Gengyun Lin, Yuxin Li, Lina Huang, Shihan Chen, Yuying Zhang, Ming Gong, Jinyu Lin, Hongbo Liang, Weichao Hao, Haisu Tao, Jian Yang","doi":"10.1016/j.acra.2026.09.031","DOIUrl":"https://doi.org/10.1016/j.acra.2026.09.031","url":null,"abstract":"<p><strong>Rationale and objectives: </strong>This study aims to develop and externally validate an interpretable machine-learning model that integrates gadolinium ethoxybenzyl diethylenetriamine pentaacetic acid magnetic resonance imaging (MRI) radiomics, qualitative MRI features, and clinical data to accurately distinguish small (≤2 cm) hepatocellular carcinomas from benign hepatic lesions.</p><p><strong>Materials and methods: </strong>This retrospective, multicenter cohort included 296 lesions (Center A: 209 [146 training, 63 internal test]; Centers B + C: 87 external validation). Three feature categories were extracted and integrated. A multistep feature selection and model tuning pipeline was performed exclusively within the training set using nested cross-validation to prevent data leakage. Seven machine learning (ML) algorithms were compared. In the external validation set, a reader study was performed to compare the diagnostic performance of the fusion model against that of radiologists. SHapley Additive exPlanations (SHAP) was employed to interpret the optimal model.</p><p><strong>Results: </strong>The fusion model combining three feature categories achieved the best diagnostic performance. In the internal test set, the fusion model yielded an area under the receiver operating characteristic curve (AUC) of 0.913, compared to 0.791 for the radiomics-only model. Its discriminative ability remained robust in an independent external validation cohort (AUC = 0.824), and remained stable after mitigating interscanner variability with ComBat harmonization (AUC = 0.806). The fusion model showed numerically higher performance than junior radiologists and performed comparably to senior radiologists. SHAP analysis identified hepatitis B virus status, alpha-fetoprotein, and specific hepatobiliary phase-derived radiomics features as key predictors.</p><p><strong>Conclusion: </strong>An interpretable ML fusion model demonstrates robust discriminative ability for evaluating small HCCs. Although its numerical advantages over the radiomics-only model and junior radiologists did not reach statistical significance because of the limited external sample size, this noninvasive tool holds significant potential as a reliable diagnostic adjunct with future dynamic threshold optimization.</p>","PeriodicalId":50928,"journal":{"name":"Academic Radiology","volume":" ","pages":""},"PeriodicalIF":4.7,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149117858","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}