The Future Direction of Artificial Intelligence in Neurosurgery

Abstract

  Artificial intelligence (AI) has become indispensable in daily life. Small devices, such as smartphones, can now efficiently perform facial recognition. Neurosurgery relies heavily on radiological imaging, making AI a promising technology for its advancement. Although the definition of AI varies among scientists, it is generally recognized as an industrial technology that simulates intelligent behavior unique to humans. The invention of deep learning (DL) is a breakthrough that has expanded AI research and is characterized by multilayer computation and the autonomous fine-tuning of parameters. Although scientists provide the ground design for DL structure, a DL device tunes and learns by itself by comparing numerous sets of input and output data. Most AI research in the field of neurosurgery can be categorized into two types : 1. qualitative assessment of lesions using radiological images solved mainly by image feature extraction algorithms, such as AlexNet, and 2. segmentation and detection of abnormal lesions within images by DL designed for semantic segmentation, such as U-Net. The qualitative assessment of radiological images of gliomas has been extensively investigated, and AI reportedly improves the diagnostic accuracy of genetic alterations in gliomas by 10% compared to those of conventional radiomic approaches. Other studies demonstrated fully automated gene alteration detection using magnetic resonance imaging (MRI) for gliomas, with apparent limitations derived from overfitting. Automated lesion detection tasks can be represented by aneurysm detection using AI-assisted MRI. A U-Net-based algorithm reportedly exhibited diagnostic accuracy with a sensitivity of 85.0% and specificity of 74.6% for automated cerebral aneurysm detection. Although this performance was inferior to that of neuroradiologists or neurosurgeons in terms of specificity (by 20%), its sensitivity was 15% higher. These results are promising for clinical applications. High sensitivity, rather than specificity, is expected in AI-assisted diagnostic technology to decrease cases of missed diagnosis by humans. Other applications include automatic glioma segmentation, which is expected to be incorporated in future clinical trials. The AI technologies in the medical field require higher accuracy and robustness than those deployed in consumer markets. These results suggest that we should view AI technologies not as magic wands but rather as sophisticated statistical algorithms with pros and cons.