Retinal Aging in 3× Tg-AD Mice Model of Alzheimer's Disease

Abstract

The retina, as part of the central nervous system (CNS), can be the perfect target for in vivo, in situ, and noninvasive neuropathology diagnosis and assessment of therapeutic efficacy. It has long been established that several age-related brain changes are more pronounced in Alzheimer's disease (AD). Nevertheless, in the retina such link is still under-explored. This study investigates the differences in the aging of the CNS through the retina of 3× Tg-AD and wild-type mice. A dedicated optical coherence tomograph imaged mice's retinas for 16 months. Two neural networks were developed to model independently each group's ages and were then applied to an independent set containing images from both groups. Our analysis shows a mean absolute error of 0.875±1.1 × 10^-2 and 1.112±1.4 × 10^-2 months, depending on training group. Our deep learning approach appears to be a reliable retinal OCT aging marker. We show that retina aging is distinct in the two classes: the presence of the three mutated human genes in the mouse genome has an impact on the aging of the retina. For mice over 4 months-old, transgenic mice consistently present a negative retina age-gap when compared to wild-type mice, regardless of training set. This appears to contradict AD observations in the brain. However, the 'black-box" nature of deep-learning implies that one cannot infer reasoning. We can only speculate that some healthy age-dependent neural adaptations may be altered in transgenic animals.

Keywords: Alzheimer's disease; age-gap; aging; animal model; artificial intelligence; deep learning; optical coherence tomography; retina.

Figure 1

Representative optical coherence tomography B-scans. Top to bottom, B-scans of a Wild type and a triple transgenic familiar Alzheimer's disease mouse model (3× Tg-AD), imaged at 1 and 16 months old, right to left, respectively.

Figure 2

Training, tuning, and testing workflow. Dataset 1 (DS1) was used for training and tuning (80/20%) of two models, M1 and M2, using only wild-type and 3× Tg-AD B-scans, respectively. Dataset 2 (DS2) containing both genotypes was used for hold-out testing with both models.

Figure 3

Kernel density estimates of the predicted age, separated per class (wild-type vs. 3× Tg-AD), for each acquisition time-point. From left to right, results for the wild-type and the 3× Tg-AD trained models. Median (dashed), and first and third quartiles are shown. *p < 0.05, **p < 0.01. Asterisk color indicates a higher median value.