OS03.7.A AUTOFLUORESCENCE BASED RECOGNITION OF BRAIN TUMORS WITH A CONVOLUTIONAL NEURAL NETWORK

Abstract

Abstract BACKGROUND The delineation of malignant brain tumors from surrounding tissue is crucial for maximal safe resection. Existing tools for intraoperative tumor visualization lack the possibility of real-time tissue classification. We therefore built a coherent fiber bundle (CFB) endoscope for autofluorescence imaging and trained a convolutional neural network (CNN) to distinguish tumor tissue (T) and non-tumor tissue (nT) in an ex vivo proof of concept study. MATERIAL AND METHODS The T-cohort included frozen tissue sections of 35 glioblastoma and 30 brain metastasis patients (10 breast cancer, 10 melanoma, 10 lung cancer). The nT-cohort comprised frozen tissue sections of hippocampi and temporal lobes from 23 patients undergoing surgery for epilepsy treatment. For image acquisition, a laser with excitation wavelength of 473 nm was coupled into a CFB and tissue autofluorescence between 500-550 nm was collected using a bandpass filter. The acquisition time of the camera was set to 700 ms. A CNN (VGG-19) was trained to differentiate the labels T from nT wherein the ground truth was defined by histopathological diagnosis. Training, validation and test sets were randomly generated across all patients in 66.6% / 16.6% / 16.6% proportion. Set generation was repeated 5 times and reported results for the independent test sets are mean values of all iterations. RESULTS In total, 3259 CFB images were acquired (T:2408, nT:851), with usually 40 CFB images per sample. Removal of CFB background was not improving classification accuracy and thus, was not performed. The classification accuracy of the CNN with a threshold probability of 0.5 for T-label assignment was 0.986 (sensitivity: 0.988, specificity: 0.982) for analysis of images and 0.98 (sensitivity: 0.971, specificity: 1) when calculating the result for each patient. To simulate clinical application where correct identification of healthy tissue is essential, we created a specificity-tuned CNN with adjusted threshold probability of 0.9 for T-label assignment. This led to a decreased accuracy of 0.945 but increased specificity on the image evaluation approach to 1 (sensitivity: 0.92). CONCLUSION Autofluorescence imaging through a CFB is capable of ex vivo recognition of brain tumor tissue with a CNN. The simplicity of the setup and minimal need for data pre-processing allows for intraoperative implementation and potentially enables real-time in vivo brain tumor delineation. To realize intraoperative transition, analysis of ex vivo bulk tissue will be upcoming to capture the signal complexity caused by tissue depth.