In this study, we utilized 2.5D fundus imaging data to develop a diagnostic framework for detecting the presence or absence of diabetic nephropathy. Since fundus images comprise multiple layers, including retinal deep blood flow and choriocapillaris structures, we implemented a 2.5D multi-instance learning approach to effectively integrate these heterogeneous image representations. The experimental phase has been completed, demonstrating the feasibility of this strategy for capturing multi-layered vascular and structural information to improve diagnostic accuracy. Fig1 shows pipeline of our study. Fig1. Workflow of this study. Methods Datasets To ensure reliable generalization, all cases were randomly allocated into training and testing cohorts at a 7:3 ratio. This process was iteratively repeated until no significant differences in clinical characteristics were observed between groups (p > 0.05). Hyperparameters were optimized through a grid-search strategy combined with five-fold cross-validation within the training set. The optimal configurations were subsequently applied to the full training cohort, and final performance was evaluated on the independent test cohort. 2.5D Deep Learning Data Preparation For each subject, multimodal CTA images were obtained, encompassing the retinal superficial plexus, retinal deep plexus, full-thickness retina, choriocapillaris, and large choroidal vessel layers. Each of these distinct image types contributed to the construction of the final training dataset. Slice-Level Model Training In slice-level experiments, each frame was treated as an independent input, and patient-level labels were weakly assigned to all frames belonging to the same subject. CNN architectures (ResNet18, ResNet50, ResNet101, DenseNet121, DenseNet201), initialized with ImageNet-pretrained weights, were employed to exploit hierarchical representations and improve domain robustness. Data Augmentation: Standardization was achieved by applying Z-score normalization across RGB channels. During training, data variability was enhanced through random cropping and horizontal/vertical flipping. Testing was limited to normalization to preserve consistency. Data Normalization: Pixel values were rescaled to [−1, 1] using min–max normalization. Cropped image patches were resized to 224 × 224 pixels via nearest-neighbor interpolation to conform to CNN input dimensions. Training Parameters: The learning rate was dynamically adjusted with a cosine decay schedule: where , , and denotes the total number of epochs. Stochastic Gradient Descent (SGD) was adopted as the optimizer, with softmax cross-entropy as the loss function. Multi-Instance Learning for Patient-Level Fusion A multi-instance learning (MIL) framework was introduced to aggregate slice-level outcomes with radiomic descriptors, thereby enhancing predictive accuracy. Two complementary strategies were developed: Predict Likelihood Histogram (PLH) and Bag of Words (BoW). In the PLH pipeline, slice-level probabilities () and predictions () from 2.5D models were binned into histograms, normalized through min–max scaling, and transformed into and feature vectors to summarize predictive distributions. Conversely, the BoW pipeline encoded unique slice-level prediction patterns as dictionary elements. Frequency counts were transformed via Term Frequency–Inverse Document Frequency (TF–IDF), generating and features that highlighted informative slice-level patterns. The final fused representation was obtained as: Feature Selection A multi-stage feature selection pipeline was applied to construct a parsimonious feature set. Initially, univariate testing retained features significantly associated with outcomes (p < 0.05). Redundancy was reduced by eliminating variables with Pearson correlation coefficients > 0.9. Subsequently, Minimum Redundancy Maximum Relevance (mRMR) selected features with high relevance to outcomes, narrowing the set to 16. Finally, Least Absolute Shrinkage and Selection Operator (LASSO) regression with L1 regularization, optimized via 10-fold cross-validation, identified the most discriminative subset. This hierarchical strategy balanced dimensionality reduction with predictive stability. MIL Signature The refined features were used to train both linear and nonlinear classifiers. Support Vector Machine (SVM) represented linear models, whereas Random Forest (RF), Extra Trees, and XGBoost captured nonlinear interactions. Hyperparameters were optimized by grid search with five-fold cross-validation. For benchmarking, the MIL -based framework was compared with traditional ensemble fusion methods. with additional details provided in Supplementary 1. Model Metrics Combined Model: Clinical characteristics were integrated with MIL-derived outputs to construct a Combined model. Evaluation Metrics: Independent test set evaluation encompassed three dimensions: discrimination, calibration, and clinical utility. Discrimination was measured using ROC-AUC with 95% confidence intervals, sensitivity, specificity, and accuracy at the Youden index threshold. Calibration was assessed by calibration curves and Hosmer–Lemeshow tests (p > 0.05 indicating good fit). Clinical benefit was evaluated with decision curve analysis (DCA) and clinical impact curves to visualize patient-level stratification. All candidate models underwent cross-validation, and the highest-performing algorithm was retained for final testing. Statistical Analysis All analyses were performed using the OnekeyAI platform (v3.5.12, Python 3.7.12). The Shapiro–Wilk test evaluated normality of continuous variables, followed by Student’s t-test or Mann–Whitney U test as appropriate. Categorical variables were compared using chi-square or Fisher’s exact tests (when expected counts < 5). No significant baseline differences were observed between cohorts (p > 0.05), confirming unbiased allocation (Table 1). Hypothesis testing and regression analyses were conducted with Statsmodels (v0.13.2), radiomic feature extraction with PyRadiomics (v3.0.1), and model training with Scikit-learn (v1.0.2). feature_name ALL train test pvalue Sex 0.844 0 96(69.06) 66(68.04) 30(71.43) 1 43(30.94) 31(31.96) 12(28.57) Table 1 Baseline characteristic of our cohorts. 以上是我分析的影像组学的方法学不分，基于以上提问要求，接着请帮我完成这份文章的方法部分，内容是跟眼部OCTA预测糖尿病肾病相关的，具体要求如下，要求写出来的内容要具有学术性，专业性，符合高分SCI杂志的要求。以中文的格式输出。 3. Results: 3.1 Patient Characteristics: 展示详细的基线资料表，并说明训练集与测试集在所有变量上均无显著差异 32。 3.2 Performance of Slice-Level Deep Learning Models: 展示Table 2 33和ROC曲线（Fig 2 34），说明ResNet101被选为最佳基础模型的原因 35。 3.3 Interpretability of the Slice-Level Model: 展示Grad-CAM可视化结果（Fig 3 36），并解读热图区域与可能的病灶关联 37。 3.4 Performance of Patient-Level MIL Signature: 展示Table 3 38和ROC曲线（Fig 4 39），证明XGBoost在测试集上的综合性能最优 40。 3.5 Comparison of Different Signatures: 展示Table 4 41和ROC曲线（Fig 5 42），核心是突出MIL模型相较于Ensemble模型的优越性 43。 3.6 Model Calibration and Clinical Utility: 展示测试集的校准曲线（Fig 6 44）和DCA曲线（Fig 7 45），证明模型的预测是可靠且具有临床净收益的 。