Towards Real-Time Confirmation of Breast Cancer in the OR Using CNN-Based Raman Spectroscopy Classification

Breast-conserving surgery is a recommended treatment for early-stage breast cancer. Recurrence and post-operative complications are potential risks when margins are not entirely removed during surgery or when timing constraints in the OR limit extensive analysis of resected tissue. Raman spectroscopy (RS), a non-destructive optical technique, enables the acquisition of molecular signatures of tissue samples allowing confirmation of different diseases, including cancer. Typically, the measured spectra must be processed and used to train conventional machine learning classifiers for cancer/normal discrimination. However, there is a lack of real-time spatially-resolved information that allows confirmation of cancer at a specific site during surgery. In this paper, we propose a tissue characterization pipeline based on convolutional neural networks (CNN), using 4 x 1D convolutional layers for automated feature extraction and a fully-connected layer as an alternative to classifying the complete RS spectra (without previous feature selection). Using 169 samples collected from 20 patients, we evaluated the performance of the proposed model, achieving an accuracy and sensitivity of 0.93(0.01) and 0.94(0.02), respectively, improving over traditional SVM-based models. Results demonstrate the potential of CNN models for classification in the OR and highlight the value of efficient signal processing to enhance their use for in-situ cancer detection.