Professor of Neurosurgery Macquarie University Sydney NSW, NSW, Australia
Disclosure(s):
Antonio Di Ieva, MD, PhD, FRACS: Abbott: Grant/Research Support (Ongoing); B Braun: Grant/Research Support (Ongoing)
Introduction: Computational Neurosurgery is a novel field where artificial intelligence (AI) is used to analyze neurosurgical diseases. We have created a framework to a) automatically segment and characterize brain tumours on MRI; b) predict the genetic subtype of gliomas on histological specimens, and c) implement AI models by means of neurosurgical expertise.
Methods: For the segmentation and classification task, we used 510 brain tumour images to train a deep learning (DL) model for automatic segmentation of the tumors’ edges and comparison of the AI-generated masks versus experts’ consensus. For the histopathology task, we digitalized 266 haematoxylin/eosin slides of gliomas (including 130 IDH-wildtype and 136 IDH-mutant, also augmented by a Generative Adversarial Network methodology) and applied a DL architecture to predict the IDH genetic status. The AI-generated data were used to implement the neuro-oncology MDT decision-making. Finally, 40 neurosurgeons were enrolled to look at brain tumors’ MRIs, and their eye tracking was recorded and analyzed to improve the extraction of salient regions of interest by the AI model.
Results: In the segmentation and classification task, we reached >90% accuracy to segment and predict tumor type. In the histopathology task, we were able to predict the genetic status with accuracy up to 95% using the DL model. The model was further tuned up by the information generated by the computational analysis of neurosurgeons’ eye tracking.
Conclusion : We have shown the robustness of AI methodology for the automatic segmentation and classification (including genetic prediction) of brain tumors on imaging. Moreover, we have also implemented a machine vision pipeline where neurosurgeons’ cognitive expertise, characterized by means of their eye gaze analysis, is used to implement AI models. Our results support the use of AI in the neuro-oncology scenario for a fast and objective computerized characterization of brain tumors and for decision-making enhancement.
