Rapid Diagnosis of Diabetic Nephropathy Using a Convolution‐Based Dual‐Branch Cross‐Modal Interaction Network

ABSTRACT Diabetic nephropathy (DN) is one of the most serious microvascular complications of diabetes. Early diagnosis is of great significance for preventing disease progression and improving patient prognosis. However, current diagnostic approaches have significant limitations in terms of sensitivity, specificity, and cost‐effectiveness. This study proposes a convolution‐based dual‐branch cross‐modal interaction network (CDCIN), which effectively integrates the complementary information from Raman and infrared spectroscopy via an innovative cross‐modal feature fusion mechanism. The core innovation of CDCIN is the construction of a multiscale cross‐modal interaction module within a dual‐branch architecture, which enables adaptive feature alignment and enhancement across the modalities and incorporates an attention mechanism to capture key biomarker information. In the comparative experiment of cross‐modal fusion methods, CDCIN achieved an accuracy of 95.79%, an AUC value of 0.9932, and an F1 score of 0.9618 for DN diagnosis. Compared with existing multimodal fusion methods such as CCFN, MMLTF, HFBSurv, and GPDBN, the accuracy was improved by 5.26%, 9.47%, 8.42%, and 4.45%, respectively. To gain deeper insights into the model's diagnostic mechanism and validate its biological plausibility, this paper analyzes the key spectral features learned by CDCIN and demonstrates that these features effectively reflect abnormal changes in key metabolites, such as phenylalanine and fatty acids, in the serum of patients with DN. This study provides a novel solution for the early screening and rapid diagnosis of DN, demonstrating significant clinical application value.