This study investigates how to improve the reliability of probability estimates produced by deep learning models for the detection of Alzheimer’s disease using MRI data. Although convolutional neural networks (CNNs) can accurately classify neurodegenerative diseases, their softmax outputs often misrepresent true classification probabilities. We evaluated four calibration methods, two parametric (logistic and probit regression) and two nonparametric (isotonic regression and Bayesian binning into quantiles), on data from 474 participants. All models improved the CNN’s calibration noticeably without reducing accuracy. Non-parametric methods achieved the best calibration results (expected calibration error ≈ 0.014 and maximum calibration error ≈ 0.025). These findings suggest that non-parametric calibration provides more reliable and clinically useful probability estimates.