Application of neural networks on in vitro-generated Raman spectra for label-free, ex vivo skeletal muscle detection

Abstract

Raman spectroscopy can detect sample molecular changes in a label-free manner but the scarce nature of precious biomedical samples and their high spectral complexity impede use of artificial intelligence (AI) for accurate analysis and diagnostics. To address this, we assessed whether in vitro-generated Raman spectra can aid AI-based neural network learning for accurate ex vivo skeletal muscle detection, as a proof-of-concept. First, a diverse in vitro Raman spectral dataset was generated followed by identification of myogenic peaks. Second, myogenic peak assignment was validated by comparison with skeletal muscle tissue. Lastly, 1-dimension (1D) convolutional neural networks (CNNs) trained from in vitro-generated Raman spectra attained 98.5% and 85.0% detection accuracy for in vitro myogenesis and ex vivo skeletal muscle, respectively. In conclusion, we demonstrated proof-of-concept for using in vitro generated data for ex vivo label-free tissue detection, which may be broadly applied to facilitate label-free Raman detection for research and clinical diagnostics.