Dual‐Branch Model‐Driven Intelligent Diagnosis of Osteoarthritis via Multidimensional Raman Spectral Data Fusion

ABSTRACT Osteoarthritis is a chronic disease, and early diagnosis is critical for clinical intervention. Raman spectroscopy is a non‐invasive and sensitive method for molecular fingerprinting and has recently proved promising for biomedical sensing and disease diagnosis. However, the peak overlap of complex biological samples often limits feature extraction accuracy. To overcome this bottleneck, we propose a novel dual‐branch convolutional neural network architecture (DBCM). By feeding one‐dimensional raw spectra and two‐dimensional spectral data simultaneously, one‐dimensional branch captures fine variations in peak intensities, and two‐dimensional branch mines nonlinear spatial correlations between frequency bins. The complementary multidimensional features effectively disentangle the overlap information. Experimentally using five‐fold cross validation results show that the DBCM provides diagnostic accuracy of 0.9142 and 0.9734 on osteoarthritis and alcohol dependence data respectively. To further improve the robustness of feature representation, we incorporate Inception module to optimize the network architecture, employing multi‐scale convolutional kernels to extract features across different frequency widths in parallel. Improved DBCM obtains accuracy of 0.9689 and F1 score of 0.9792 on the osteoarthritis dataset. Our results suggest that DBCM deeply integrates multidimensional Raman features which provides a high precision tool for early osteoarthritis diagnosis and a new model paradigm for deep learning applications in biomedical spectral analysis.