Efficient Net- B0 baseline for three class Alzheimer MRI with imbalance remedies

Class imbalance in structural MRI can bias Alzheimer disease classifiers toward majority categories and raise the risk of clinically costly false negatives. We compare two simple and widely used remedies within the same EfficientNet-B0 transfer learning pipeline for three class slice level classification of AD, CN, and MCI. Method A uses balanced sampling each epoch with standard cross entropy. Method B uses natural sampling with class weighted cross entropy. With identical preprocessing, augmentations, label smoothing 0.05, the AdamW optimizer, and a two-phase schedule with 8 epochs of warm up followed by 10 epochs of fine tuning, we evaluate on a stratified 15% validation split that reflects natural prevalence with AD 18,440 slices, CN 26,892, and MCI 42,744. Method B attains accuracy 0.9938, macro F1 0.9938, and AD recall 0.9946, exceeding Method A with 0.9839, 0.9829, and 0.9646. AD false negatives decreased from 98% of method A to 15% of method B. The gain arises from exposure to all unique training samples each epoch and from an objective that matches the unbalanced validation distribution by reweighting errors without altering batch priors. These results support class weighted loss as a strong and architecture agnostic baseline for imbalanced AD MRI and agree with prior evidence that weighting can outperform synthetic oversampling on related Alzheimer datasets.