A lightweight deep learning pipeline for MRI-based differential diagnosis of Alzheimer’s disease and mild cognitive impairment

Abstract Background Alzheimer’s disease (AD) accounts for 60–70% of dementia cases, with mild cognitive impairment (MCI) as a prodromal stage. Early diagnosis is critical. This study aimed to develop and validate a lightweight deep learning pipeline for automated classification of AD, MCI, and cognitively normal (CN) using routine 1.5 T MRI. Method We analyzed 567 ADNI participants (95 CN, 368 MCI, 104 AD ). Preprocessing of T1-weighted MRI included skull stripping, resampling, and intensity normalization. Eleven consecutive coronal slices covering the hippocampus were extracted per subject. Two CNNs (ConvNeXt-Tiny, EfficientNet-B0) were trained on a stratified 80/10/10 patient-level split. Binary classification tasks: AD vs CN and AD vs MCI. Result ConvNeXt-Tiny achieved AUCs of 0.98 (AD vs. CN) and 0.74 (AD vs. MCI), outperforming EfficientNet-B0 (0.98 and 0.66, respectively). Sensitivity/specificity for AD vs CN were 0.94/0.96; for AD vs MCI, 0.68/0.71. Conclusion CNN-based classification of 1.5 T MRI was feasible for AD vs CN, but performance dropped substantially in AD vs MCI, which suggested a need for more discriminative features or multimodal input. The lightweight pipeline was suitable for integration into routine neuroimaging practice.