gms | German Medical Science

Objective: Reliable intraoperative tumor classification remains a challenge as conventional fast pathology is limited by spatial and time factors. Confocal laser endomicroscopy (CLE) with cellular resolution can be used intraoperatively during fluorescence-guided brain tumor surgery in real-time and will be available for in vivo use. To overcome observer dependent and sampling errors we established a pattern recognition and high image volume based automated approach. We aimed to predict tumor type by applying a residual convolutional network to image data obtained by CLE.

Methods: Human brain tumor specimen were achieved intraoperatively from 26 patients (9 metastasis, 8 glioblastoma, 9 meningioma). Fresh tissue was used for further investigation immediately. The samples were stained with fluorescein in vitro and analyzed with CLE. Tissue samples were acquired for comparison with routine histological analysis. We trained a residual convolutional neural network for all three tumor types (metastasis, glioblastoma, meningioma) and built a predictive level for the outputs with confidence intervals. To increase the size of training set and prevent overfitting, we augmented the training set images by introducing random rotations, translations and flips. We trained the network with a triplet loss, which allowed the network to compare images. This gave us the ability to receive a confidence value for predictions.

Results: Multiple CLE-images were obtained from each patient with a total number of 11164 images. With our neural network model we achieved a ratio of correct predicted label of 81%, which was confirmed with a second validation set of 79%, without definition of confidence. Applying a confidence rate of 95%, the prediction accuracies were increased to 99,5% in the training set and 93% in the validation set, with 5464 and 185 images used, respectively. The second fold provided similar results using a validation set with a total of 1576 images. Here, the predictive level was 92,6%, with 190 images classified presuming a confidence rate above 95%.

Conclusion: We developed a model that allows fast, reliable and automated analysis of images acquired by fluorescein based CLE. We thereby have a tool that enables CLE to be used on the fly during brain tumor surgery and facilitates intraoperative decision-making. Further in vivo studies are required for advanced differentiation of pathologies and normal as well as reactive brain tissue to assess the status CLE can have in the neurosurgical workflow.