To develop and test, in two distinctive datasets, a pretrained gated fusion model that predicts CNS amyloid-β (aβ) positivity from multi-contrast three-dimensional (3D) brain MRI in cognitively normal individuals.

CNS aβ is quantified with PET, which is costly and restricted worldwide. Identifying 3D MRI-based surrogates could enable earlier Alzheimer’s screening and guide  serologic/PET testing.

To extract representations from T1-weighted and T2-FLAIR and to address limited labeled datasets, we pretrained two convolutional neural networks using self-supervised contrastive learning on 989 unlabeled T1-weighted and 693 FLAIR scans from ADNI. We separately trained a gated fusion network (GFN) to adaptively weight the MRI modalities, and the fused representations were passed to a multilayer network to predict aβ-positivity. A new ADNI dataset included 279 cognitively normal participants, 92 aβ-positive (AV45 SUVR>1.11), and 187 aβ-negative: 90% were used for training  (4-fold cross-validation) and 10% were held-out for testing. We validated on a single-institution dataset (68 cognitively normal; 20 Aβ-positive and 48 Aβ-negative; 28 APOEε4 carriers). We report mean(±SD) performance metrics and comparative performances in APOEε4 carriers vs non-carriers.

On the ADNI testing set, the model achieved an accuracy of 0.63±0.04, sensitivity 0.52±0.03, specificity 0.73±0.11, PPV 0.67±0.08 and NPV 0.60±0.03. The GFN assigned greater weights to T1-weighted compared to FLAIR. On our small institutional set, sensitivity was 0.51±0.22, specificity 0.67±0.13, PPV 0.39±0.03 and NPV 0.78±0.04. The model had significantly greater PPV (0.56 vs 0.25, p=0.03) and lower NPV (0.53 vs 0.90, p=0.03) in APOEε4 carriers compared to non-carriers.

Predicting aβ-positivity in cognitively normal using 3D-MRI remains suboptimal, with better PPV in APOEε4 carriers. We leveraged self-supervised contrastive pretraining to reduce dependence on large labeled cohorts and support scalable model development. Additionally, our GFN enabled adaptive weighting of various MRI modalities, allowing handling missing modalities and future integration of relevant clinical data.