Optimal Raman Spectral Classifcation Model Based on Differentiable Architecture Search of Hybrid Structure Network for Disease Diagnosis

Abstract

Identification and classification are important application areas of surface-enhanced Raman spectroscopy (SERS). Substance is identified via the chemical finger-print function of Raman spectroscopy. Diseases can be diagnosed through biofluidic Raman spectrum analysis and classification accordingly. Since bio-fluidic, such as serumurineand tissue fluid contains various substances, Raman spectrum is too complex to be classified manually. The optimization of deep learning classification model is critical in diagnosis accuracy improvement. Here we propose, for the first, applying DARSHN algorithm in automatic diagnosis model design and optimization. DARSHN was applied to serialize the discrete search space. Optimal structural solution was generated through approximate gradient descent subsequently. This research suggested that DARSHN can be used in the optimization of classification models automatically and effectively. Its advantages in the application of serum SERS-based cancer diagnosis compared to residual network spectral classification models were shown in this paper.