An Integrated Raman Platform with Embedded AI for Intraoperative Real-Time Cancer Detection

Raman spectroscopy workflows are often fragmented across proprietary tools and ad hoc data-processing scripts, which slow decision-making and limit reproducibility and auditability. We present an integrated Raman platform with a Python-based GUI application that interfaces with the optical system(s) to unify instrument control, automate data acquisition, noise removal, signal processing, and run an embedded AI model for real-time classification of experimental samples. The system automatically archives intermediate and final outputs for auditability. We validated the platform's output for two types of experimental specimens: Tylenol and peritumoral biological samples from human laryngeal tissues. We compared the results with other studies, commercial software, and previous data-processing algorithms. The results obtained were consistent across all comparators. To demonstrate native analytics within the same workflow, we also developed and embedded a 1D convolutional neural network (CNN) tailored for biological Raman spectra. The multilayered CNN was trained on ex vivo human laryngeal tissue Raman spectra and was evaluated across 50 independent runs. The model achieved a mean test accuracy of 0.8929 ± 0.0213, a sensitivity of 0.9086 ± 0.0321, a specificity of 0.8698 ± 0.0434, and a mean AUC of 0.9506 ± 0.0142. The average latency across 50 runs was 2.82 s from signal acquisition to prediction, with device acquisition accounting for most of the time. The key innovation is not only the embedded AI but also an end-to-end, auditable workflow that unifies device control, acquisition, signal processing, visualization, and the archival of intermediate and final outputs, with optional real-time inference within a single platform.