Zhan Feng

Rationale and objectives: To evaluate the performance of four large language models (LLMs) for automated feature extraction and clear cell likelihood score version 2.0 (ccLS v2.0) assignment from free-text reports using a multi-step prompting strategy. Materials and methods: This retrospective multicenter study included 1,048 magnetic resonance imaging (MRI) reports of renal masses from three institutions (2020–2026). Four LLMs (DeepSeek-V3.2, Qwen 3.5, GPT-5.3, and Gemini 3.0) were evaluated for automated ccLS v2.0 assignment through a three-stage prompting workflow: (1) feature extraction, (2) rule-based criteria matching, and (3) final categorization. Performance was evaluated against expert radiologist consensus-based ccLS v2.0 assignments. Accordingly, these results primarily reflect agreement in task execution rather than diagnostic accuracy for ccRCC. Results: Gemini 3.0 achieved the highest overall accuracy, correctly assigning ccLS categories in 683 of 802 reports in the internal cohort (85.2%; 95% CI, 82.5%–87.5%) and 211 of 246 reports in the external validation cohort (85.8%; 95% CI, 80.9%–89.6%). All LLMs achieved accuracy greater than 80% for assigning high-risk categories (ccLS 4–5). Gemini 3.0 showed comparatively strong performance in extracting individual imaging features, such as T2-weighted imaging (T2WI) hyperintensity. However, all models showed reduced performance for features requiring multi-step interpretation, particularly the arterial-to-delayed enhancement ratio (ADR), with Qwen 3.5 showing the weakest performance for these complex features. In the pathology subgroup of the internal cohort ( n = 237), expert report-based scores yielded an AUC of 0.799 for identifying ccRCC, with LLM-assigned scores showing AUCs of 0.768–0.791 and small AUC differences from the expert standard (−0.031 to −0.008) with overlapping 95% CIs. Conclusion: LLMs employing multi-step prompting, particularly Gemini 3.0, demonstrated favorable performance in assigning ccLS v2.0 categories from free-text reports. Despite limitations in complex feature reasoning, these models showed potential for supporting automated ccLS assessment in radiology workflows; however, in their current implementation, they should be considered decision-support tools that require expert verification rather than autonomous, unsupervised scoring.

Purpose Accurate risk stratification for patients with high-risk, non-metastatic clear cell renal cell carcinoma (ccRCC) remains a clinical challenge. Our aim is to develop and validate an imaging model based on intratumoral habitat imaging for non-invasive assessment of progression-free survival (PFS) in this high-interest population. Methods A retrospective analysis was performed on a cohort of 220 patients with ccRCC from three hospitals. Preoperative CT images and clinical data were comprehensively analyzed. K-mean clustering method was used to generate habitat subregions using contrast-enhanced CT images. Radiomic features were extracted from habitat subregions and the entire tumor, Random Survival Forest for PFS prediction were constructed using conventional radiomics features, habitat subregions-derived radiomics. Model evaluation included calibration curve, receiver operating characteristic curve (ROC) and Kaplan-Meier survival analysis. Results The Habitat 3 model demonstrated favorable prognostic performance. In the training cohort, 1-, 3-, and 5-year AUCs were 0.92, 0.92, and 0.78, respectively. In the external test set, the model maintained robustness with 1-, 3-, and 5-year AUCs of 0.87, 0.74, and 0.94, outperforming traditional radiomics and other Habitat models. Furthermore, patients stratified by Habitat 3 showed significant survival differences between high-risk and low-risk groups in both cohorts. Conclusion We proposed a habitat-based imaging technique for the preliminary analysis of high-risk ccRCC. This retrospective study suggests that habitat signatures are associated with PFS, serving as a potential non-invasive complement to provide insights into tumor heterogeneity.

Background: The impact of diabetes on the management and outcomes of patients with borderline CT-derived fractional flow reserve (FFR_CT) remains unclear. Methods: This multicenter study enrolled symptomatic patients with suspected coronary artery disease who underwent Coronary computed tomography angiography (CCTA) between June 2021 and May 2023, yielding FFR_CT values between 0.70 and 0.80. Revascularization occurring within 90 days after CCTA was documented. The endpoint was major adverse cardiovascular events (MACE), as a composite of all-cause death, nonfatal myocardial infarction, and unplanned revascularization. Outcomes were analyzed using Cox proportional hazards models, while the relationship between FFR_CT and MACE was examined using restricted cubic spline analysis (RCS). Results: This analysis included 1515 patients with borderline FFR_CT values, comprising 503 (33.2%) with diabetes. Over a median follow-up of 985 days, 117 MACE occurred. Multivariate analysis showed that revascularization was independently associated with a reduced risk of the endpoint, a protective effect consistent in both non-diabetic (adjusted HR [aHR] 0.53, 95% CI 0.29–0.96; p = 0.036) and diabetic patients (aHR 0.25, 95% CI 0.09–0.71; p = 0.009). RCS revealed a significant non-linear relationship between FFR_CT and MACE in non-diabetic patients (p = 0.002). Conclusions: In patients with borderline FFR_CT, revascularization was linked to a lower incidence of MACE, and this association was consistent regardless of diabetes status.

Objectives To assess the feasibility and accuracy of using deep learning to generate simulated contrast-enhanced T1-weighted rectal MRI scans from pre-contrast MRI sequences in rectal cancer patients. Methods This study included 514 patients with pathologically confirmed rectal carcinoma who underwent contrast-enhanced MRI at three academic institutions. A two-dimensional generator adversarial network was used to simulate contrast-enhanced MRI scans from pre-contrast T1-weighted images. Quantitative assessments of image similarity, error metrics, and Dice coefficient for tumor overlap were performed. Three radiologists, blinded to the contrast method, evaluated image quality, tumor enhancement, and extramural vascular invasion (EMVI) in 104 paired real and simulated scans. Tumor size correlation was analyzed with intraclass correlation coefficients (ICCs), and modified Bland-Altman plots were used to assess agreement. ROC curves evaluated diagnostic performance of EMVI detection, and quality scores were compared using the McNemar test. Results Simulated scans demonstrated high similarity to real scans (structural similarity index: 0.82, Dice coefficient: 0.86 ± 0.14 for tumor overlap). Tumor size measurements correlated strongly (ICC: 0.76–0.88), with minimal differences in length (0.11 mm) and depth (0.10 mm). ROC analysis revealed slightly higher AUC values for real scans (0.702, 0.772, 0.713) compared to simulated scans (0.685, 0.694, 0.661) for three radiologists, respectively. Nearly all synthetic scans (98%) were rated diagnostically suitable. Conclusion Deep learning-generated simulated contrast-enhanced MRI scans demonstrated comparability to real scans in terms of tumor size, enhancement area, and image quality, suggesting their potential as a supportive tool for diagnosis and treatment evaluation.

Background To investigate the prognostic significance of visceral fat to predict recurrence-free survival in patients with localized clear cell renal cell carcinoma (ccRCC). Methods This study included patients with localized ccRCC who had undergone curative surgery. Visceral, subcutaneous fat and the relative visceral fat area (rVFA) were quantified utilizing preoperative CT images. The association between rVFA and recurrence-free survival (RFS) was explored using restricted cubic splines, Cox proportional hazards regression, and Kaplan–Meier survival analysis. The predictive capability of rVFA was evaluated using the Boruta algorithm, a predictive model was developed using the random survival forest (RSF) with results visualized via Shapley additive explanations (SHAP). Results Four hundred and forty six patients were included. The follow-up median duration was 48.5 months, during which 57 patients experienced metastasis or death. Restricted cubic spline model revealed a nonlinear association between rVFA and tumor progression, exhibiting a U-shaped curve trend with an inflection point at 0.40. Beyond this threshold, rVFA was significantly correlated with an increased risk of progression. The RSF model yielded an area under the curve (AUC) of 0.89 for predicting 1-year RFS, 0.73 for 3-year RFS, and 0.75 for 5-year RFS. Both Boruta and SHAP analyses identified rVFA as a significant predictive feature. Conclusions A U-shaped association between rVFA and risk of tumor recurrence was observed among patients with ccRCC. A high rVFA is significantly correlated with an increased risk of adverse events, thus indicating that rVFA is a potential indicator for predicting the prognosis of ccRCC.

Purpose: To investigate the association between CT radiomics-based tumor heterogeneity and recurrence-free survival (RFS) in high-risk clear cell renal cell carcinoma (ccRCC) after nephrectomy, and to determine whether integrating CT radiomics with clinicopathological model enhances recurrence risk prediction for adjuvant treatment decisions. Methods: This retrospective study included 194 patients with high-risk ccRCC undergoing nephrectomy. A radiomics model based on random survival forest was developed in the training set, using radiomics features extracted from pre-operative corticomedullary phase images. The performance of radiomics, Leibovich score, and the combined model were evaluated using Kaplan-Meier survival analysis, time-dependent receiver operating characteristic curves (time-AUC), time-dependent Brier scores, and decision curve analysis in external test set. Results: During follow-up, 62 patients experienced recurrence. The radiomics model demonstrated superior predictive performance compared to the Leibovich score, with higher time-dependent AUCs (1-year: 0.882 vs. 0.781; 2-year: 0.865 vs. 0.762; 3-year: 0.793 vs. 0.797; all p < 0.05) and better calibration (lower Brier scores) in the test set. Decision curve analysis demonstrated that the combined model provided the highest net benefit, particularly for 2- to 3-year recurrence risk predictions. Conclusions: For high-risk ccRCC, CT radiomics provides incremental prognostic value beyond conventional clinicopathological models, enabling more precise recurrence risk stratification. This approach bridges imaging and precision oncology, with potential to optimize surveillance protocols and adjuvant therapy trial design.

Purpose: To develop and validate an accurate computed tomography–based radiomics model for predicting high-grade (micropapillary/solid) patterns in T1-stage lung invasive adenocarcinoma (IAC) after propensity score matching (PSM). Materials and Methods: We enrolled 546 participants from 2 cohorts with histologically diagnosed lung IAC after complete surgical resection between January 2020 and August 2021. The patients were divided into high-grade and non–high-grade groups and matched using PSM. Matched patient HRCT images were used to delineate regions of interest from tumors and extract radiomics features, and the random forest method was used to construct a radiomics model. The area under the receiver operating characteristic curve (area under the curve) was used to evaluate the model’s performance, and external validation was performed to assess the model’s generalizability. Results: Before PSM, there was no statistically significant difference in age between the two groups, though nodule type and sex exhibited significant differences (P < 0.05) in both cohorts. After PSM, we matched 176 and 97 pairs of patients in the 2 cohorts. In both cohorts, sex and nodule type were equal between the two groups, with a higher percentage of males and solid nodules in both groups. Our model exhibited moderate predictive performance after PSM, with area under the curve values of 0.75 (95% CI: 0.70-0.80) and 0.71 (95% CI: 0.63-0.80) for the development and external validation cohorts, respectively. Conclusion: Although the nodule type compromised the validity of the model’s performance, our results suggest that our acute computed tomography–based radiomics model could preoperatively predict micropapillary/solid patterns in patients with stage I lung IAC after PSM.

Objective: To exploit the improved prediction performance based on dynamic contrast-enhanced (DCE) MRI by using dynamic radiomics for microvascular invasion (MVI) in hepatocellular carcinoma (HCC). Methods: We retrospectively included 175 and 75 HCC patients who underwent preoperative DCE-MRI from September 2019 to August 2022 in institution 1 (development cohort) and institution 2 (validation cohort), respectively. Static radiomics features were extracted from the mask, arterial, portal venous, and equilibrium phase images and used to construct dynamic features. The static, dynamic, and dynamic–static radiomics (SR, DR, and DSR) signatures were separately constructed based on the feature selection method of LASSO and classification algorithm of logistic regression. The receiver operating characteristic (ROC) curves and the area under the curve (AUC) were plotted to evaluate and compare the predictive performance of each signature. Results: In the three radiomics signatures, the DSR signature performed the best. The AUCs of the SR, DR, and DSR signatures in the training set were 0.750, 0.751 and 0.805, respectively, while in the external validation set, the corresponding AUCs were 0.706, 0756 and 0.777. The DSR signature showed significant improvement over the SR signature in predicting MVI status (training cohort: P = 0.019; validation cohort: P = 0.044). After external validation, the AUC value of the SR signature decreased from 0.750 to 0.706, while the AUC value of the DR signature did not show a decline (AUCs: 0.756 vs. 0.751). Conclusions: The dynamic radiomics had an improved effect on the MVI prediction in HCC, compared with the static DCE MRI-based radiomics models.