[3J1-GS-1] Fundamental AI, theory

[3J1-GS-1-04] Reliability in AI for Brain Imaging Analysis

〇Atsushi Kawaguchi¹, Yuko Ishimaru¹, Yuka Takao¹, Ryosuke Osako¹ (1. Saga University)

Keywords:Brain Imaging Analysis, AI, Biostatistics

In this paper, after reviewing and discussing the reliability of brain image analysis, we conducted an experiment to evaluate the explainable AI analysis methods on real brain image data. The review claimed that brain imaging analyses that yield new findings need to be evaluated for reliability even in small sample size situations, while a large sample size is necessary to provide sufficient reliability in the analysis. We evaluated the explainable AI analysis methods on the classification problem of brain differences between Alzheimer's disease and normal groups for the OASIS data. The experimental result showed that in the case of evaluation using the F1 score, the GradCAM++ method for the ResNet model best matched the true regions.