2018 AANS Annual Scientific Meeting-508. Convolutional Neural Networks Provide Rapid Intraoperative Diagnosis Of Neurosurgical Specimens Imaged With Stimulated Raman Histology

508. Convolutional Neural Networks Provide Rapid Intraoperative Diagnosis Of Neurosurgical Specimens Imaged With Stimulated Raman Histology

Back to course

Video Transcription

Video Summary

Dr. Todd Holland discusses the use of convolutional neural networks (CNNs) and simulated Raman histology (SRH) in rapid interoperative diagnosis of neurosurgical specimens. The traditional pathology workflow involves freezing, cutting, mounting, and staining specimens, which can limit interpretation and delay care. In contrast, SRH is a label-free, tissue-processing-free imaging technique. CNNs, specifically the Google Inception V3 CNN, can interpret SRH images and provide a diagnosis in a fraction of the time. The CNN is trained on a large dataset of SRH images and achieves an accuracy of 75.6% on small tile images. When integrated at a patient level, the CNN achieves 95% accuracy. The classification of tumor classes and important decision points in the operating room are highly accurate, and the method can also detect infiltrating tumor margins. The combination of artificial intelligence and SRH shows promise in improving brain tumor surgery outcomes. Dr. Todd Holland gives credit to his mentors, Dr. Daniel Orringer, Dr. Corinne Marasteanu, and Dr. Cormack Maher, for their support.

Asset Caption

Todd Hollon, MD

Keywords

convolutional neural networks

simulated Raman histology

neurosurgical specimens

rapid interoperative diagnosis

label-free imaging