Objective:To investigate the value of a deep learning radiomics (DLR) nomogram model based on intra-tumoral and peri-tumoral MRI features for differentiating IgG4-related ophthalmic disease (IgG4-ROD) from orbital mucosa-associated lymphoid tissue (MALT) lymphoma.Methods:This was a case-control study. The clinical and imaging data of 233 patients pathologically confirmed with either IgG4-ROD or orbital MALT lymphoma were retrospective collected between January 2020 and December 2024 from the Second Hospital of Jilin University (Center 1) and Beijing Tongren Hospital, Capital Medical University (Center 2). Patients from Center 1 ( n=158) were used as the training cohort, while those from Center 2 ( n=75) served as the validation cohort. Among the cases, 102 were IgG4-ROD (70 in training, 32 in validation) and 131 were orbital MALT lymphoma (88 in training, 43 in validation). Univariate and multivariate logistic regression analyses were used to identify independent clinical imaging predictors and build a clinical imaging model. Based on T 1WI, T 2WI, and diffusion weighted images, intra-tumoral regions were manually delineated, a 2 mm peri-tumoral margin was automatically generated, and both regions were combined as a single region of interest for radiomics feature extraction. Deep learning features were extracted using a ResNet-50 backbone, and after feature selection and dimensionality reduction, a DLR model was constructed. The clinical imaging features and DLR features were integrated to build a combined nomogram model. Model performance in differentiating IgG4-ROD from orbital MALT lymphoma was assessed using receiver operating characteristic curves, calibration curves, and decision curve analysis. The area under the curve (AUC) were compared using the DeLong test. Results:Bilateral orbital involvement ( OR=1.983, 95% CI 1.166-2.843, P=0.046) and extraocular muscle involvement ( OR=1.246, 95% CI 1.079-1.764, P=0.015) were identified as independent predictors for distinguishing IgG4-ROD from orbital MALT lymphoma and were used to construct the clinical model. Fourteen features (9 radiomics and 5 deep learning features) were selected for the DLR model, and a nomogram was developed. In the training set, the AUCs for the clinical model, DLR model, and nomogram were 0.762 (95% CI 0.712-0.812), 0.865 (95% CI 0.822-0.908), and 0.943 (95% CI 0.909-0.953), respectively. In the validation set, the AUCs were 0.733 (95% CI 0.675-0.791), 0.823 (95% CI 0.762-0.884), and 0.924 (95% CI 0.902-0.958), respectively. The nomogram showed significantly higher AUCs than those of the clinical and DLR models alone (training set: Z=3.92, 2.87, P0.001, P=0.004; validation set: Z=3.25, 2.46, P=0.001, 0.014). Calibration curves indicated good agreement between predicted and actual IgG4-ROD incidence, and decision curve analysis demonstrated the highest net benefit for the nomogram. Conclusion:A nomogram that incorporates both intra-tumoral and peri-tumoral DLR features and clinical imaging characteristics demonstrates excellent performance in distinguishing IgG4-ROD from orbital MALT lymphoma.

Objective:To investigate the value of a deep learning radiomics (DLR) nomogram model based on intra-tumoral and peri-tumoral MRI features for differentiating IgG4-related ophthalmic disease (IgG4-ROD) from orbital mucosa-associated lymphoid tissue (MALT) lymphoma.Methods:This was a case-control study. The clinical and imaging data of 233 patients pathologically confirmed with either IgG4-ROD or orbital MALT lymphoma were retrospective collected between January 2020 and December 2024 from the Second Hospital of Jilin University (Center 1) and Beijing Tongren Hospital, Capital Medical University (Center 2). Patients from Center 1 ( n=158) were used as the training cohort, while those from Center 2 ( n=75) served as the validation cohort. Among the cases, 102 were IgG4-ROD (70 in training, 32 in validation) and 131 were orbital MALT lymphoma (88 in training, 43 in validation). Univariate and multivariate logistic regression analyses were used to identify independent clinical imaging predictors and build a clinical imaging model. Based on T 1WI, T 2WI, and diffusion weighted images, intra-tumoral regions were manually delineated, a 2 mm peri-tumoral margin was automatically generated, and both regions were combined as a single region of interest for radiomics feature extraction. Deep learning features were extracted using a ResNet-50 backbone, and after feature selection and dimensionality reduction, a DLR model was constructed. The clinical imaging features and DLR features were integrated to build a combined nomogram model. Model performance in differentiating IgG4-ROD from orbital MALT lymphoma was assessed using receiver operating characteristic curves, calibration curves, and decision curve analysis. The area under the curve (AUC) were compared using the DeLong test. Results:Bilateral orbital involvement ( OR=1.983, 95% CI 1.166-2.843, P=0.046) and extraocular muscle involvement ( OR=1.246, 95% CI 1.079-1.764, P=0.015) were identified as independent predictors for distinguishing IgG4-ROD from orbital MALT lymphoma and were used to construct the clinical model. Fourteen features (9 radiomics and 5 deep learning features) were selected for the DLR model, and a nomogram was developed. In the training set, the AUCs for the clinical model, DLR model, and nomogram were 0.762 (95% CI 0.712-0.812), 0.865 (95% CI 0.822-0.908), and 0.943 (95% CI 0.909-0.953), respectively. In the validation set, the AUCs were 0.733 (95% CI 0.675-0.791), 0.823 (95% CI 0.762-0.884), and 0.924 (95% CI 0.902-0.958), respectively. The nomogram showed significantly higher AUCs than those of the clinical and DLR models alone (training set: Z=3.92, 2.87, P0.001, P=0.004; validation set: Z=3.25, 2.46, P=0.001, 0.014). Calibration curves indicated good agreement between predicted and actual IgG4-ROD incidence, and decision curve analysis demonstrated the highest net benefit for the nomogram. Conclusion:A nomogram that incorporates both intra-tumoral and peri-tumoral DLR features and clinical imaging characteristics demonstrates excellent performance in distinguishing IgG4-ROD from orbital MALT lymphoma.