Introduction: Amyloid pathology is a defining hallmark of Alzheimer’s disease(AD). Yet conventional amyloid detection methods remainexpensive, invasive, and limited in accessibility—restrictingearly detection and large-scale screening. The current studyvalidates an algorithm that synthesizes amyloid-pathology images using routine MRI scans. This approach offersa scalable, cost-effective, and non-invasive method to detectamyloid pathology that could facilitate early Alzheimer'sscreening and clinical phenotyping.

Methods: A deep convolutional neural network was trained tosynthesize amyloid pathology images from T1-weighted MRIscans. The model was evaluated on 192 participants   acrossthe AD spectrum. We compared SUVrs (standardized uptakevalue ratios) calculated using the whole cerebellum as thereference region from AV-45 PET images and the amyloidpathology images. Amyloid positivity was defined usingboth, gold standard PET SUVr > 1.11 (Landau et al., 2021)and CSF tTau/Aβ42 > 0.28 (Amft et al., 2022) as groundtruths, independently. A comparison of SUVrs from the PETand amyloid pathology images was performedusing correlation, equivalence testing, and classificationmetrics. 

Results: The amyloid pathology image-derived SUVrs showed significant correlation with gold standard PET SUVrs (r = 0.66, p < 0.0001)and a mean absolute percentage difference of 10.16%. Equivalence testing confirmed statistical equivalence (95% CI: 0.0057to 0.0477, p < 0.001). Using amyloid PET SUVr as ground truth, the model achieved an accuracy of 82%, sensitivity of 73%,specificity of 88%, and AUROC of 85%. While using CSF derived ground truth, the model achieved 81% accuracy, 82%sensitivity, 80% specificity and 81% AUROC. Subgroup analysis revealed the greatest absolute SUVr difference in MCI,followed by AD, and the least in cognitively normal (CN) participants. 

Conclusion: This proof-of-concept study demonstrates thefeasibility of extracting amyloid pathology signalfrom routine brain MRI. SUVrs extracted from bothimages demonstrate significant alignment withboth PET- and CSF-derived ground truths, showingpotential as a reliable screening tool. Thesefindings suggest that it is possible to deriveclinically meaningful inferences about Alzheimer’spathology without the need for invasive orexpensive procedures, paving the way for scalableand accessible early detection solutions. 

Clinical Relevance and Application: 

• Enables non-invasive, cost-effective and accessible amyloidpathology detection from routine MRI.

• Leverages standard clinical neuroimaging to support scalable,cost-effective screening, especially in low-resource settings.

• Facilitates early intervention through screening, therapyeligibility assessment and disease phenotyping.

• Supports clinical decision making by enabling longitudinalmonitoring and informing Alzheimer’s disease managementstrategies.