(P 245) IN VIVO USE OF RAMAN SPECTROSCOPY IN SOFT TISSUE SARCOMA RESECTION

Objective: Negative-margin resection is one of the most significant prognostic factors for retroperitoneal soft tissue sarcoma (STS) survival. Raman spectroscopy is an optical analysis technique based on interaction between fixed-wavelength photons and tissue samples. It provides information on human tissue samples at a molecular level, in a non-destructive manner, and sufficiently quickly not to disrupt surgical workflow. Neoplastic transformation of a healthy tissue involves important biomolecular changes that generate significant changes in the tissue’s corresponding Raman spectrum. Machine learning classification algorithms can detect these spectrum modifications to classify a given sample as tumoral or healthy tissue. This study aims to determine the sensitivity, specificity, and accuracy of a Raman spectroscopy handheld point probe paired with a random forest machine learning classification algorithm in detecting soft tissue sarcoma in an in vivo environment during STS resection.

Methods: Patients were recruited amongst those referred to Maisonneuve-Rosemont Hospital (Montreal, QC) from January 2021 to June 2023 for retroperitoneal and extremity STS resection. Raman spectra were acquired during STS resection using a Raman handheld point probe (UltraProbe), developed in collaboration between the Laboratory of Radiological Optics (LRO) from Polytechnique Montréal and EmVision LLC. Spectra were acquired from healthy adipose, muscular, and fibrotic tissue in vivo. Spectra were acquired from the sarcoma specimen following resection. For every spectrum acquired, a tissue sample was resected and examined by an expert pathologist to establish a histopathological diagnosis. Every spectrum was added to a database along with its associated histopathological diagnosis. Using this database, sequential binary machine learning random forest classification algorithms were developed and trained to predict the diagnosis associated with new Raman spectra: soft tissue sarcoma or healthy tissue. Spectra classification was done in two steps, the first being to perform a binary classification to predict whether the spectrum of interest has a high lipid content or not using a random forest machine learning classification algorithm. A second classification followed to classify high-lipid content Raman spectra as healthy adipose tissue or liposarcoma. For spectra classified as non lipidic, a multiclass classifier was built to classify as healthy tissue or other sarcoma subtypes. Sensitivity, specificity, and accuracy calculations were made to compare the classifier’s prediction to the histopathological diagnosis used as a gold-standard.

Results: 29 patients have been recruited in the study. Preliminary results following complete histopathological diagnosis and spectra processing are available for 9 participants. The preliminary dataset includes 332 labeled Raman spectra. Well differentiated liposarcoma was the only sarcoma subtype with sufficient spectra for classification. The classification of Raman spectra between lipidic and non-lipidic tissue was performed with a sensitivity of 97%, a specificity of 91%, and an accuracy of 95%. The second classification of lipidic spectra as well differentiated liposarcoma or healthy adipose tissue was performed with a sensitivity of 93%, a specificity of 95%, and an accuracy of 94%.

Conclusion: The presented results are from a preliminary dataset. Full results will be available in November 2023. The presented results show that Raman spectroscopy can be used to detect well differentiated liposarcomas in an in vivo setting. Its sensitivity, specificity and accuracy were greater than 90%. Raman spectroscopy is an optical tissue analysis technique that is fast, precise, non-destructive, and can thus be repeated as needed during sarcoma resection surgery without disrupting the surgical workflow. Further studies will be needed to determine if its use can lead to a reduced rate of microscopically positive margin resections.