Multiscale Convolutional Neural Network of Raman Spectra of Human Serum for Hepatitis B Disease Diagnosis

In this study, we proposed a multiscale convolutional neural network (MsCNN) that can screen the Raman spectra of the hepatitis B (HB) serum rapidly without baseline correction. First, the Raman spectra were measured in the serums of 435 patients diagnosed with a HB virus (HBV) infection and 499 patients with non-HBV infections. The analysis showed that the Raman spectra of the serums were significantly different in the range of 400-3000 cm(-1) between HB patients and non-HB patients. Then, the MsCNN model was used to extract the nonlinear features from coarse to fine in the Raman spectrum. Finally, extracted fine-grained features were placed into the fully connected layer for classification. The results demonstrated that the accuracy, sensitivity, and specificity of the MsCNN model are 97.86%, 98.94%, and 96.79%, respectively, without baseline correction. Compared to the traditional machine learning method, the model achieved the highest classification accuracy on the HB data set. Therefore, multiscale convolutional neural network provides an effective technical means for Raman spectroscopy of the HBV serum.