ObjectivesTo investigate the predictive value for diabetic kidney disease (DKD) by utilizing deep learning to automatically segment fundus panoramic images and constructing multidimensional radiomics models integrated with clinical parameters. MethodsA diagnostic trial study. From December 2022 to March 2024, 353 patients with type 2 diabetes mellitus who were admitted for the first time to Department of Endocrinology of Yangzhou University Affiliated Jiangdu People's Hospital were included in the study. Among them, 114 cases had diabetic kidney disease (DKD), and 239 cases were non-DKD patients. All patients underwent non-mydriatic color fundus photography of both eyes, capturing 45° field of view fundus images centered on the macula (panoramic fundus images). A pre-trained U-Net model was used to automatically segment regions of interest (ROI) in the fundus panoramic images, and batch processing was performed to segment ROI for all images. Radiomics features were extracted separately from the ROI of both left and right fundus panoramic images. Feature-level fusion (feature concatenation) was applied to merge the radiomics features from both eyes, resulting in a fused feature set. Features from the left eye, right eye, and the fused set underwent feature selection using t-tests, Pearson correlation analysis, and LASSO regression to identify the optimal features for model building. Finally, the dataset was partitioned into a training set and an independent validation set at a 7:3 ratio. The training set underwent 5-fold cross-validation for hyperparameter tuning, ultimately yielding the optimal algorithm model classifier, separate radiomics models were built for the left eye, right eye, and the feature-level fusion set. Additionally, a decision-level fusion model (ensemble voting) was constructed by combining the outputs (results) of the left and right eye models. A clinical parameter model was also built based on multivariate analysis results. The area under the receiver operating characteristic curve (AUC) was used as the primary quantitative evaluation metric. DeLong's test compared AUC differences between models. The net reclassification index (NRI) and decision curve analysis (DCA) were employed to assess the superiority between different models. ResultsThe results of the ROC analysis showed that in the training and validation sets, the AUC values for the clinical model, left-eye radiomics model, right-eye radiomics model, feature-level fusion model, and decision-level fusion model were 0.826, 0.847, 0.883, 0.890, 0.907 and 0.588, 0.646, 0.642, 0.657, 0.689, respectively. The results of the DeLong test showed that in the training set, the AUC of the decision-level fusion model was significantly higher than that of the clinical model and the left-eye model (P=0.010,＜0.001). The AUC of the feature-level fusion model was significantly higher than that of the left-eye model (P=0.020). However, in the validation set, no statistically significant differences in AUC were observed among the models (P＞0.05). The results of the NRI Analysis showed that compared to the clinical model, the NRI values for all four radiomics models were positive in both training and validation sets, indicating superior DKD prediction performance by the radiomics models. Compared to the decision-level fusion model, the NRI values for the left-eye, right-eye, and feature-level fusion models were negative in both sets, suggesting that the decision-level fusion model had the best performance. The results of the DCA analysis showed that in both training and validation sets, the decision-level fusion model provided greater net clinical benefit across a range of threshold probabilities compared to the other four models. ConclusionThe radiomics model based on automatically segmented panoramic fundus images can predict the risk of DKD occurrence, with the integrated model of both eyes demonstrating higher predictive performance.
