gms | German Medical Science

Objective: Accurate intraoperative tissue diagnostics could impact on decision making regarding the extend of resection during brain tumor surgery. Stimulated Raman histology (SRH) is a label-free optical imaging method that uses different biochemical properties of tissue to generate a hematoxylin-eosin-like image and, in combination with an artificial intelligence-based image classifier, offers the opportunity to obtain rapid intraoperative tissue diagnosis. The goal of this study was to report on our initial experience with SRH to evaluate its accuracy in comparison to final tissue diagnosis.

Methods: We included 70 consecutive adult cases with brain tumors that underwent a surgical resection in a single academic institution. Tissue samples were taken intraoperatively and examined using the NIO laser imaging system (Invenio-Imaging^®) classifying each tissue specimen to 1 out of 14 predefined tissue classes according to a probability score. We compared results of the SRH classifier with the highest probability score to the respective final histopathological result. Prediction accuracy was evaluated by logistic regression and Receiver Operator Curve (ROC) analysis.

Results: According to final tissue diagnosis, we included 19 meningiomas (27%), 17 gliomas (24%), 6 pituitary adenomas (9%), 9 metastasis (13%) and 19 other tumor entities such as neurinoma or ependymoma (27%; total n=70). Regarding accuracy of intraoperative SRH predictions, regression analysis showed an Area Under the Curve (AUC) suggesting agreement of predictions with final diagnosis in 76% (95% C.I. 0.64-0.89, p=0.0008). Looking at each tumor entity the results varied. The highest accuracy was obtained for meningiomas (AUC=0.95; 95% C.I. 0.85-1, p=0.015), followed by glioma (AUC=0.88; 95% C.I. 0.67-1, p=0.09) and pituitary adenoma (AUC=0.71; 95% C.I. 0.26-1, p=0.38). The accuracy for metastases (AUC=0.70; 95% C.I. 0.43-0.98, p=0,19).

Conclusion: Our initial experience with SRH shows that this novel imaging option is a promising approach to obtain rapid intraoperative tissue diagnosis. However, refinement of artificial intelligence-based automated image classification has to be driven forward to improve prediction accuracy and to improve reliability. This could enable to individually tailor the desired extent of resection to the proposed tumor entity to improve surgical and oncological outcomes.

Figure 1 [Fig. 1]