초록: Mild Cognitive Impairment (MCI) is characterized by measurable declines in one or more cognitive domains while individuals largely maintain independence in daily life. Early diagnosis is critical, as 10R11;15% of MCI patients progress to dementia annually, making timely interventionsR12;such as pharmacological treatments, cognitive training, and lifestyle changesR12;essential for delaying or preventing further decline. Despite the importance of early detection, accurate diagnosis of MCI often requires expensive neuroimaging and a battery of neuropsychological tests, due to the heterogeneous nature of its symptoms and progression. Consequently, relying solely on brain imaging is often insufficient.
In this study, we apply deep learning algorithms using a multimodal approach to examine how diagnostic performance differs when using imaging data alone versus combining it with neuropsychological test results. Finally, after achieving accurate diagnostic predictions, we employ integrated gradient-based saliency maps to identify the brain regions and neuropsychological features that most significantly contribute to the model’s decision-making, and we discuss the implications of these findings.
Keywords: multimodality, deep learning, MCI, interpretability
2025-1학기 전산전자공학대학원 여섯 번째 세미나 강의를 소개합니다.
Mild Cognitive Impairment (MCI) is characterized by measurable declines in one or more cognitive domains while individuals largely maintain independence in daily life. Early diagnosis is critical, as 10R11;15% of MCI patients progress to dementia annually, making timely interventionsR12;such as pharmacological treatments, cognitive training, and lifestyle changesR12;essential for delaying or preventing further decline. Despite the importance of early detection, accurate diagnosis of MCI often requires expensive neuroimaging and a battery of neuropsychological tests, due to the heterogeneous nature of its symptoms and progression. Consequently, relying solely on brain imaging is often insufficient.
In this study, we apply deep learning algorithms using a multimodal approach to examine how diagnostic performance differs when using imaging data alone versus combining it with neuropsychological test results. Finally, after achieving accurate diagnostic predictions, we employ integrated gradient-based saliency maps to identify the brain regions and neuropsychological features that most significantly contribute to the model’s decision-making, and we discuss the implications of these findings.
Keywords: multimodality, deep learning, MCI, interpretability
본 세미나는 오프라인(영어)로 진행됩니다. 학부생, 대학원생, 교직원 누구나 참여 가능하니 많은 관심과 참여 바랍니다.