This was a retrospective, single-center observational cross-sectional study. We retrospectively reviewed patients pathologically diagnosed with OAML or IgG4-ROD at the Department of Ophthalmology, West China Hospital, Sichuan University, between May 2010 and April 2024.

Clinical data were extracted from electronic and paper medical records, including demographic characteristics, presenting symptoms, lesion location, best-corrected visual acuity (BCVA), serum IgG4 levels, systemic involvement, treatment strategies, pathologic findings, and clinical outcomes. Lesion sites were determined based on imaging findings and intraoperative observations.

Corticosteroid exposure was defined as intravenous, oral, or periocular administration within 1 month prior to surgery, excluding topical formulations. Follow-up was conducted through scheduled outpatient visits and telephone interviews. Study endpoints included recurrence, progression-free survival (PFS), overall survival (OS), systemic dissemination, and treatment-related complications.

For OAML, relapse or progression was defined as disease advancement per the Ann Arbor–Cotswolds staging system or clinical worsening of the primary lesion.

For IgG4-ROD, relapse was defined as an increase of > 2 points in the IgG4-RD Response Index (IgG4-RD RI) after initial remission.

This study employed imaging-based and pathology-based artificial intelligence models for differentiating OAML from IgG4-ROD. Pathology whole-slide images and radiological CT/MRI images were processed using standardized preprocessing pipelines. Detailed procedures for pathology whole slide imaging (WSI) tiling, radiological image normalization, and dataset partitioning are provided in Supplementary Methods S1–S3. All pathology and radiology data were de-identified prior to analysis. To prevent information leakage, all experiments were conducted with strict patient-level separation between training, validation, and test sets. All image slices derived from the same patient were assigned to the same dataset partition.

Technical details are summarized below, with full methodological descriptions provided in the Supplementary Methods. Figure 1 illustrates the overall frameworks of the proposed models. All models are grounded in prototype learning, which represents each class by a semantic prototype in the embedding space. This design facilitates structured representation learning under weak supervision and class imbalance while enhancing model robustness and interpretability.

Overview of 3 frameworks for differential diagnosis between IgG4-ROD and OAML. (A) The PCMIL framework for pathological differential diagnosis. Whole-slide images are partitioned into 256 × 256 patches and encoded using an ImageNet-pretrained ResNet-50 backbone to extract instance-level embeddings. Attention-based multiple instance learning aggregates patch features into slide-level representations. Class prototypes are introduced to regularize feature geometry by promoting intra-class compactness and inter-class separation; (B) The PCMIL framework for radiological differential diagnosis. Radiological images are encoded using a ResNet-50 backbone to produce feature vectors. A dual-balanced representation learning strategy is adopted, incorporating feature-level prototype regularization and inter-class balancing with frequency-aware logit scaling. Final predictions are obtained through a cosine similarity–based prototype classifier. (C) The Pathology-Informed Prototype Prompting (PP) framework. Pathological prototype banks derived from the PCMIL pathology model are transferred to the radiological model to provide disease-specific semantic guidance. Parallel cross-attention enables alignment between radiological and pathological prototypes, generating prompt-informed radiological representations for improved differentiation between IgG4-ROD and OAML.

For pathological image analysis, we developed a prototype-clustering multiple instance learning framework (PCMIL) that aggregates patch-level features into slide-level predictions and encourages intra-class compactness and inter-class separability through prototype regularization. For radiological image analysis, we proposed a dual-balanced representation network (DBRNet) that harmonizes feature geometry and classifier responses to mitigate class imbalance and potential distributional shifts.

To incorporate pathological information into radiological learning, we introduced a pathology-informed training strategy based on prototype prompting (PP). This approach enables cross-modal knowledge transfer during model development while preserving unimodal radiological inference at test time. To reduce overfitting under moderate sample size, backbone feature extractors were frozen where appropriate, early stopping was applied based on validation performance, and model complexity was controlled through prototype number regularization.

Detailed mathematical formulations, loss functions, and optimization strategies for PCMIL, DBRNet, and prototype prompting are provided in Supplementary Methods S4–S5.

Model performance was evaluated using the AUC, accuracy, precision, specificity, sensitivity, and F1 score. AUC was used as the primary metric to assess discriminative performance under class imbalance. Specificity and sensitivity were reported to reflect clinical trade-offs between false-positive and false-negative predictions. Comparative experiments were conducted against representative multiple instance learning models for pathology and commonly used convolutional and transformer-based backbones for radiology. Descriptions of baseline models and comparative experimental settings are summarized in Supplementary Methods S6. Model selection was performed exclusively on the validation set, and all reported results were obtained from an independent held-out test set.

Missing data for clinical and laboratory variables were imputed using the missForest algorithm. Missingness was < 10% for all variables included in the analysis. Categorical and continuous variables were compared using chi-square and Mann–Whitney U tests, respectively. For IgG4-ROD, univariate and multivariate regression identified predictors of serum IgG4 concentration, reported as coefficients with SEs and p values, and visualized in coefficient plots. Survival was analyzed by Cox proportional hazards models, with Kaplan–Meier curves illustrating group differences. Binary logistic regression distinguished OAML from IgG4-ROD, and diagnostic performance was assessed by ROC curves, AUC, sensitivity, specificity, and calibration. All analyses were performed in R (v4.4.1; R Foundation, Vienna, Austria) with statistical significance set at P <.05. Data preprocessing and statistical analyses were implemented in Python v3.12.11, using NumPy v1.26.4, SciPy v1.13.0, pandas v2.2.2, SimpleITK v2.5.2, scikit-learn v1.5.0, and Matplotlib v3.9.0. All deep learning models were developed and trained using PyTorch v2.9.0 + cu126. Because of the retrospective design and real-world data availability, sample sizes varied across analyses (e.g., imaging, serology, immunohistochemistry, and molecular testing). All analyses were conducted on all eligible cases with available data for the corresponding modality, and no imputation was applied for imaging or pathological outcomes.

During the study period, 357 OAML and 148 IgG4-ROD patients were pathologically diagnosed after surgical excision of orbital or ocular adnexal masses. After exclusion of cases with incomplete documentation or inadequate follow-up, 287 OAML and 113 IgG4-ROD patients were included in the final analysis.

The OAML group (n = 287) consisted of 171 men and 116 women, with a mean age of 58.3 ± 13.4 years. Mean best-corrected visual acuity (BCVA) was 0.71 ± 0.34, and the median symptom-to-diagnosis interval was 12 months (range 1–369). Most lesions were unilateral (96.2%).

The IgG4-ROD group (n = 113) included 69 men and 44 women, with a mean age of 49.8 ± 12.9 years. Mean BCVA was 0.83 ± 0.29, and the median symptom-to-diagnosis interval was 15 months (range 1–147). Bilateral involvement was observed in 49.6% of cases. Among these patients, 105 had direct immunohistochemical confirmation, and 8 initially labeled as benign lymphoid hyperplasia were reclassified after review.

Baseline characteristics are summarized in Table 1 .