Discriminate Raman/GFP Spectra of Yeast Mitochondria Using Convolutional Neural Network

Raman spectroscopy is one of the well-known techniques that be able to identify molecular species in the complex system. In the present work, 633 nm Raman laser and GFP excitation light are irradiated at the sample simultaneously. It allows us to know whether Raman spectra are indeed taken from mitochondria. However, the mixture spectra of RamanlGFP are very complicated and impossible to correctly identify manually. Therefore, convolutional neural network is integrated to our work in order to distinguish RamanlGFP spectra with and without signal from GFP as GFP-positive and GFP-negative spectra, respectively. The system architecture is the adapted visual geometry group (VGG) with 14 convolutional layers. Batch normalization is added, and fully-connected layer is increased to 5 layers. Various size of RamanlGFP images were used to train our VGG network. The black and white background color also provided to gain the best result. The highest accuracy of 87.42% was achieved to separate GFP-positive and -negative spectra with black background and 68⨯52 pixel images. In the final step, the separated GFP-positive and -negative spectra were distributed in two different folders for further Raman spectra analysis of mitochondria.