ScholarWorks@성균관대학교

초록

Mild Cognitive Impairment represents a clinically significant transitional condition in the progression toward Alzheimer’s Disease, underscoring the need for diagnostic tools that are both accurate and clinically reliable. This study presents a deep learning based diagnostic framework designed to support trustworthy medical decision making by explicitly addressing three pillars of trustworthy artificial intelligence, including adversarial robustness, fairness across gender groups and visual explainability. The proposed approach analyzes gray matter, white matter, and cerebrospinal fluid through dedicated tissue segmentation, followed by an ensemble of deep learning classifiers optimized using Bayesian optimization. The framework was evaluated on the Alzheimer’s Disease Neuroimaging Initiative cohort and demonstrated strong and consistent diagnostic performance, outperforming representative baseline methods in distinguishing progressive and stable forms of Mild Cognitive Impairment. The generalization capability was further assessed using an independent external dataset. To ensure clinical relevance, visual explainability of disease related brain regions and model robustness were jointly evaluated by a medical domain expert, supporting a clinician in the loop workflow and demonstrating the practical utility of the system in real clinical settings. The results indicate that the proposed framework provides an interpretable, robust, and fair diagnostic solution that aligns with emerging expectations for trustworthiness in medical imaging and offers meaningful support for clinical decision making.

키워드