Introduction: Understanding the histopathology primary spine tumors and their proximity/involvement with surrounding normal structures strongly influences the extent of surgical resection. Wide surgical margins are associated with better outcomes after spine tumor resection, though they may be difficult to achieve at times due to visualization and confidence in tissue classification as pathological. Current methods for intraoperative tissue identification are limited, leading to potential incomplete resection or injury to surrounding structures. Additionally, pathologic classification requires tissue resection and a time consuming, effort-intensive frozen and permanent pathology process. We have developed a novel technology, the “TumorID”, which uses endogenous fluorescence spectroscopy paired with machine learning (ML) methods to predict tumor pathology prior to resection and distinguish normal muscle, bone, neural structures, and tumor. The TumorID requires no exogenous dyes and is nondestructive.
Methods: We used TumorID to scan ex vivo tissue specimens (normal bone, muscle, and tumor) from 12 patients and in vivo intra-operative tissue from 5 additional patients with primary or metastatic spine tumors as our training and validation cohorts, respectively. We are currently scanning 5 additional in vivo patients as a prospective testing cohort. Utilizing these training and validation data, we conducted rigorous ML analysis with seven different algorithms: multi-layer perceptron, random forest, XGBoost, support vector classifier, naive Bayes, k-nearest neighbors, and logistic regression. Predictions were made on every individual scanned point on the tissue. ML models were tuned using 5-fold nested cross validation and the final model was selected by maximizing Area Under the Receiver Operating Curve (AUROC).
Results: Our training and validation patient cohorts had various pathologies including metastasis from various primary cancers, chordoma, meningioma, solitary fibrous tumor, giant cell tumor, and osteosarcoma. Our test cohort is being prospectively collected, with final test metrics to be reported. Previous work demonstrates high predictive power.
Conclusion : Utilizing ML, the TumorID demonstrated feasibility to, in under one second, predict the pathology of a tumor and differentiate it from normal tissue in vivo. We hope this technology will eventually allow for more quantitative, confident surgical decision making and safer, wider margins for spine tumor patients.