Use of artificial intelligence in neuro-ophthalmological diagnosis: learning from real data and development of methods to address current and future challenges

It is important to have a critical view of the support provided by Artificial Intelligence (AI) in medical context, in order to trust this support. The objective was to measure/compare unidimensional uncertainty of an AI and a human performing the same task by a cross-sectional study. It was given to a simple algorithm written in Python (blob detection, OpenCV) and to an ophthalmologist the task of detecting a two-dimensional pattern (center of the optical disc) in 1,000 digital images of normal/abnormal fundoscopies. Algorithm performed the task 1x, human performed the task 2x, both using digital register of spatial coordinates. Machine's unidimensional level of uncertainty was measured by the respective comparison of the x and y coordinates recorded by machine and human. Human's unidimensional level of uncertainty was measured by comparing the coordinates recorded by human itself. Data analysis was performed using R AI failed to detect the target pattern onlyin two images. On average, man and machine showed a higher level of uncertainty in the ycoordinates, which was greater (~100 units) in machine's performance. The measure of uncertainty of AI and humans in the same task can help understand AI limitations and define its usefulness as a medical support tool.