Deep learning-enabled soft tissue tumor localization using spatially offset Raman spectral analysis: in-silico investigations

Abstract

Raman spectroscopy and its derivatives have gained wide acceptance among optical biopsy tools for tissue discrimination. However, the identification and localization of subsurface soft tissue tumors are still challenging. Several designs for the Raman probe have been proposed to this effect, among which spatially offset Raman spectroscopy (SORS) could offer a potential solution. This paper attempts to demonstrate the simultaneous identification of subsurface adenoma depth and thickness using Convolutional Neural Networks applied on Monte Carlo simulated SORS signals. The application of transfer learning model resulted in a better root mean square error (RMSE) of 4.40% for depth prediction as compared to the 7%–25% RMSE demonstrated by previous reports. Simultaneous thickness prediction is demonstrated for the first time with 8.42% RMSE.