Ex-vivo Raman spectroscopy and AI-based classification of soft tissue sarcomas

Abstract

Soft tissue sarcomas (STS) are a diverse and rare group of malignant tumors arising from the connective tissues of the body, including fibrous tissue, muscles, fat, nerves, and blood vessels. The heterogeneity and infrequency of these tumors pose significant challenges in both diagnosis and treatment. Surgical resection remains the primary treatment strategy, often complemented by radiation or chemotherapy, contingent upon the tumor’s size, location, and stage. However, current methods for assessing intraoperative margins are limited, underscoring the need for improved approaches that enhance both efficiency and accuracy. This study investigates the potential of microscopic Raman spectroscopy for distinguishing between different subtypes of soft tissue sarcomas, benign tumors, and normal tissue. Ex-vivo Raman measurements were conducted using a 633 nm excitation wavelength on samples obtained from surgical resections of seven patients (286,672 spectra). After pre-processing of the data, a custom ResNet architecture was developed to accurately classify the different tissue types, achieving an overall weighted accuracy of 97.1% and a clinical alert rate of 1.46%, a critical metric for quantifying the misclassification of malignant tissues. These findings suggest that single Raman spectra could serve as a rapid, non-invasive tool for surgical guidance, aiding in the precise identification of abnormal tissue types and margins.