Deep learning-based fiberoptic Raman spectroscopy improves in vivo diagnosis of nasopharyngeal carcinoma at endoscopy (Conference Presentation)

The development of rapid and objective diagnostic techniques with high accuracy is highly desirable for real-time in vivo cancer diagnosis and characterization during endoscopic examination. This work reports a deep learning-based fiberoptic Raman technique for improving in vivo cancer detection of nasopharyngeal carcinoma (NPC) in clinical settings. We have developed a robust cancer diagnostic platform based on deep neural network (DNN) model in combination with fiberoptic Raman endoscopic technique for effectively extracting latent discriminative features contained in in vivo tissue Raman spectra. We applied the platform onto the tasks of predicting new NPC patients as well as follow-up of post-irradiated patients at endoscopy. A better diagnostic performance was achieved in the testing dataset by using this diagnostic platform as compared to the classic chemometric classification methods such as partial least squares-discriminate analysis (PLSDA). This work demonstrates that DNN-based fiberoptic Raman technique is more effective and reliable for NPC classification, particularly robust for clinical prediction of new NPC patients and post-irradiated patients surveillance.