HAI Book 2025 - Flipbook - Page 414
Lyu, Xueying
103
Amyloid-Tau to neurodegeneration mismatch from inter-modality
image translation using deep learning
Xueying Lyu1, Paul Yushkevich2, Mengjin Dong3, Pulkit Khandelwal4, Christopher Brown5,
Michael Duong6, Yue Li7, Long Xie8, Laura Wisse9, Sandhitsu Das10, David Wolk11
1
University of Pennsylvania, Philadelphia, PA, US
University of Pennsylvania, Philadelphia, PA, US
3
University of Pennsylvania, Philadelphia, PA, US
4
University of Pennsylvania, Philadelphia, PA, US
5
University of Pennsylvania, Philadelphia, PA, US
6
University of Pennsylvania, Philadelphia, PA, US
7
University of Pennsylvania, Philadelphia, PA, US
8
University of Pennsylvania, Philadelphia, PA, US
9
Lund University, Lund, SE
10
University of Pennsylvania, Philadelphia, PA, US
11
University of Pennsylvania, Philadelphia, PA, US
2
Background: Prior work demonstrated the mismatch between tau pathology (T) and neurodegeneration (N) can
reveal co-pathologies (e.g., vascular disease, TDP-43) and resilience to Alzheimer9s disease (AD). Tau is
considered the driver of neurodegeneration, but its interaction with amyloid (A) may jointly contribute to brain
atrophy in a spatially complex manner. This study advanced exploration of AD pathology and neurodegeneration
mismatch by creating synthetic images of N based on both T and A using inter-modality image translation,
allowing a neural network to model non-linear, spatial relationships between tau, amyloid, and
neurodegeneration. Predicted images were compared to actual images to quantify mismatch.
Method: We studied 184 amyloid-positive (A+) symptomatic patients from ADNI, each with paired Tau PET and MRI
scans representing T and N, alongside Centiloid values for A. We trained a 3D-UNet to learn AT-N relationship by
synthesizing cortical thickness map from Tau SUVR and Centiloid-derived images (Figure 1). The reconstruction
error (predicted versus actual thickness) across 100 regions of interest (ROIs) was used to quantify the AT-N
mismatch.
HAI2025 - 414
