Spectroscopy and Machine Learning Based Rapid Point-of-Care Assessment of Core Needle Cancer Biopsies

Abstract

AbstractSolid tumor needle biopsies are essential to confirm malignancy and assess for actionable characteristics or genetic alterations to guide treatment selection. Ensuring that sufficient and suitable material is acquired for tumor profiling, while minimizing patient risk, remains a critical unmet need. Here, we evaluated the performance characteristics of transmission optical spectroscopy for rapid identification of malignant tissue in core needle biopsies (CNB). Human kidney biopsy specimens (545 CNB from 102 patients, 5583 spectra for analysis) were analyzed directly on core biopsy needles with a custom-built optical spectroscopy instrument. Machine learning classifiers were trained to differentiate malignant from normal tissue spectra. Classifiers were compared using receiver operating characteristics analysis and sensitivity and specificity were calculated relative to a histopathologic gold standard. The best performing algorithm was the random forest (sensitivity 96% and 93%, specificity 90% and 93% at the level of individual spectra and full CNB, respectively).Ex-vivospectroscopy paired with machine learning paves the way towards rapid and accurate characterization of CNB at the time of tissue acquisition and improving tumor biopsy quality.