NIMG-33. MULTICENTER, PROSPECTIVE VALIDATION OF AUTOMATED INTRAOPERATIVE NEUROPATHOLOGY USING STIMULATED RAMAN HISTOLOGY AND CONVOLUTIONAL NEURAL NETWORKS. (5th November 2018)

• Record Type: Journal Article
• Authors: Hollon, Todd; Pandian, Balaji; Heth, Jason; Sagher, Oren; Maher, Cormac; Sullivan, Steve; Garton, Hugh; Thompson, Greg; Save, Akshay; Marie, Tamara; Boyett, Deborah; Petridis, Petros; McKhann, Guy; Muraszko, Karin; Bruce, Jeffrey; Camelo-Piragua, Sandra; Canoll, Peter; Orringer, Daniel
• Is Part Of: Neuro-oncology. Volume 20(2018)Supplement 6
• Page Start: vi183
• Page End: vi183
• Publication Date: 2018-11-05
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noy148.759
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: INTRODUCTION: Accurate intraoperative diagnosis is essential for providing optimal neurosurgical care. In many centers caring for brain tumor patients, neuropathology resources are limited. To augment existing neuropathology resources, we developed and validated a new paradigm combining optical histology and artificial intelligence (AI) to accurately predict diagnosis during brain tumor surgery. METHODS: A total of 1026 specimens from 501 patients undergoing brain tumor resection at two tertiary hospitals were imaged using an optical technique, called stimulated Raman histology (SRH). SRH images were used to train and validate a convolutional neural network (CNN) for state-of-the-art computer vision. We redesigned the GoogleNet InceptionV3 CNN architecture to optimize performance on SRH histologic image fields of view (FOVs) and trained the network using 466 patients (3.1 million unique 300m2 FOVs) to classify into 13 common brain tumor subtypes. Final intraoperative diagnosis was determined using the most commonly predicted FOV diagnosis within each specimen. Model testing was completed on 1 million unique FOVs from 35 prospectively enrolled patients whose data was not included in the training set. RESULTS: In the validation set, our trained CNN differentiated lesional from normal tissue with 100% accuracy, surgical from nonsurgical lesions with 100% accuracy, glial from non-glial tumors with 100% accuracy. When evaluating our model for tumor subtype classification, we achieved an accuracy of 97% (35/36 patients) compared to final clinical diagnosis. Corresponding clinical frozen section diagnostic accuracy was 97% and interrater agreement between CNN and clinical frozen section diagnosis was near-perfect (k>0.96). The sole CNN error was misclassification of a WHO grade 1 pilocytic astrocytoma as a WHO grade II astrocytoma. CONCLUSION: Our prospective, multi-institutional validation suggests that AI can be applied to predict diagnosis in neurosurgical specimens in an automated fashion. AI-based diagnosis may ultimately be used to augment the current neuropathology workflow where resources are limited.