145: THE USE OF MACHINE LEARNING FOR REAL-TIME DETECTION OF OESOPHAGEAL AND GASTRIC CANCER BASED ON DIFFUSE REFLECTANCE SPECTROSCOPY: A VALIDATION STUDY

Abstract

Abstract Background and aim The lack of a fast and accurate intraoperative tumour margin assessment tool contributes to high positive circumferential resection margin rates for oesophageal and gastric cancers, which is associated with local recurrence and poor long-term survival. Diffuse reflectance spectroscopy (DRS) can provide non-invasive, accurate, and real-time tissue classification based on quantification of light reflectance in tissue, providing a unique optical fingerprint of the tissue. This study aimed to validate the use of DRS for discrimination between normal and cancerous tissues based on ex-vivo gastric and oesophageal cancer resection specimens. Methods Consecutive patients undergoing resections for gastric and oesophageal carcinomas at a tertiary referral centre in London between July 2020 and June 2021 were included. Reflectance data in the 468-720 nm spectral range were recorded on ex-vivo specimens within 15 minutes of surgical specimen excision. Based on video recordings of the data acquisition procedure and probe tip tracking, sampling sites were correlated with histological reports manually labeled as either normal or cancer (Figure 1). Following spectral data normalisation and feature selection, four supervised machine learning approaches were tested for classification using 5-fold cross-validation. Overall classification accuracy and Area Under the Curve (AUC) across machine learning approaches were compared. Results A total of 32 patients (median age 68, 75% male) were included. 11,862 mean spectra (2990 spectra for normal oesophagus, 4628 for normal stomach, 2305 for gastric cancer, and 1939 for oesophageal cancer) were collected. For oesophageal cancer, classification accuracy was 96.22 ± 0.50 and AUC was 99.24 ± 0.19, while for gastric cancer, the accuracy was 93.86 ± 0.66, and AUC 98.50 ± 0.28 for gastric cancer. For both malignancy types, Extreme Gradient Boosting was the best performing classifier. The differences in spectral data between patients receiving neoadjuvant treatment and treatment naïve were non-significant. Conclusion Machine learning can be used to accurately differentiate between normal and cancerous oesophageal and gastric tissues in an ex-vivo setting, based on DRS data. Features used for class discrimination should be investigated to find possible physiological correlates of spectral data differences. Future intraoperative validation is required to assess the utility of this technology for real-time tumour margin mapping and its effect on long-term outcomes.