BACKGROUND: Current neuroimaging frameworks,

including the NIA-AA 2024 Revised Criteria, describe a

"plateau" in amyloid-beta (A-Beta) burden during the

symptomatic stages of Alzheimer's disease. While PET

imaging provides a vital macroscopic view, its resolution may

mask ongoing microscopic accumulation. We developed a

computational pathology pipeline to segment and classify

various A-Beta burden, using high-resolution histological

quantification to unmask continuous pathological growth

across Thal phases.

METHODS: We assembled a corpus of 7,940

neuropathologist-validated amyloid lesion instances

sampled from 321 whole-slide images stained for Aβ (6E10

and 6F3D), spanning five neuroanatomical regions (middle

frontal, inferior parietal, visual, parahippocampal, and motor

cortex). A two-stage pipeline was developed: (1) an

EfficientNet-B4 U-Net segments Aβ-immunoreactive deposits

to quantify pixel-level total burden, and (2) a classifier built

on UNI2-h, a pathology foundation model, adapted with a

trainable convolutional branch through gated cross-attention

(ViT-Adapter) and deployed as a dual-specialist, five-fold

cross-validated ensemble, assigns each deposit to one of five

classes: coarse-grained plaque (CGP), classic cored plaque

(CCP), capillary (CAA1) and parenchymal/leptomeningeal

(CAA2) cerebral amyloid angiopathy, or non-specific/other.

Internal performance was assessed by out-of-fold (OOF)

cross-validation; external validity was assessed by correlating

model-derived per-region deposit counts with independent

neuropathologist semi-quantitative scores.

RESULTS: The classifier achieved an OOF macro-F1 of

0.92 (per-class F1: 0.91 CGP, 0.90 CCP, 0.92 CAA1, 0.95 CAA2,

0.91 non-specific). Model-derived per-region deposit counts

correlated with independent neuropathologist semiquantitative

scores, with strongest agreement in the region

where each morphology predominates: CGP in middle

frontal cortex (Spearman ρ=0.755), CAA2 across all regions

(ρ=0.64–0.75), CAA1 in visual cortex (ρ=0.742), and CCP

across regions (ρ=0.55–0.67); all p<0.005.

CONCLUSIONS: A pathology foundation-model

classifier, coupled to deep segmentation and refined through

iterative neuropathologist quality control, enables accurate,

reproducible, morphology-resolved quantification of cerebral

Aβ pathology across regions, with high concordance to expert

scoring. This provides a scalable, higher-fidelity tool for

precision neuropathological staging and validating antiamyloid

therapeutic response.