Deep learning analysis of urine-derived stem cell mitochondrial morphology as a non-invasive Alzheimer's disease biomarker.

Alzheimer's disease (AD), closely associated with mitochondrial dysfunction, currently lacks convenient and non-invasive biomarkers for mitochondrial assessment. In this study, we developed an artificial intelligence framework leveraging live urine-derived stem cell (USC) mitochondrial fluorescence imaging to investigate differences between cognitively impaired individuals (AD and mild cognitive impairment (MCI)) and cognitively normal (CN) subjects. Mitochondrial fluorescence images from living HeLa cells were first segmented, and two binary classification models based on the ResNet-18 convolutional neural network were trained to identify mitochondrial hyperfission and hyperfusion relative to normal morphology. The models demonstrated robust performance in detecting intermediate mitochondrial states during validation. When applied to USCs, the system effectively distinguished mitochondrial patterns associated with cognitive impairment, highlighting its potential for the early detection of Alzheimer's disease and merits further validation in larger, independent cohorts.