Diagnostic Accuracy of a Confocal Laser Endomicroscope for In Vivo Differentiation Between Normal Injured And Tumor Tissue During Fluorescein-Guided Glioma Resection: Laboratory Investigation

Abstract

Glioma resection with fluorescein sodium (FNa) guidance has a potential drawback of nonspecific leakage of FNa from nontumor areas with a compromised blood-brain barrier. We investigated the diagnostic accuracy of invivo confocal laser endomicroscopy (CLE) after FNa administration to differentiate normal brain, injured normal brain, and tumor tissue in an animal glioma model. GL261-Luc2 gliomas in C57BL/6 mice were used as a brain tumor model. CLE images of normal, injured normal, and tumor brain tissues were collected after intravenous FNa administration. Correlative sections stained with hematoxylin and eosin were taken at the same sites. A set of 40 CLE images was given to 1 neuropathologist and 3 neurosurgeons to assess diagnostic accuracy and rate image quality (1-10 scale). Additionally, we developed a deep convolution neural network (DCNN) model for automatic image classification. The mean observer accuracy for correct diagnosis of glioma compared with either injured or uninjured brain using CLE images was 85%, and the DCNN model accuracy was 80%. For differentiation of tumor from nontumor tissue, the experts' mean accuracy, specificity, and sensitivity were 90%, 86%, and 96%, respectively, with high interobserver agreement overall (Cohen κ= 0.74). The percentage of correctly identified images was significantly higher for images with a quality rating >5 (104/116, 90%) than for images with a quality rating ≤5 (32/44, 73%) (P=0.007). With sufficient FNa present in tissues, CLE was an effective tool for intraoperative differentiation among normal, injured normal, and tumor brain tissue. Clinical studies are warranted to confirm these findings.