Real and synthetic multimodal retinal images for the diagnosis of Alzheimer’s Disease with convolutional neural networks

SLOOTWEG, Ivan R.

Real and synthetic multimodal retinal images for the diagnosis of Alzheimer’s Disease with convolutional neural networks

Files

Slootweg, I. s-1001424MSc-MKI92-Thesis-2024.pdf (28.85 MB)

Authors

SLOOTWEG, Ivan R.

Issue Date

2024-06-01

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/18872

Abstract

Abnormalities in the retinal nerve fiber layers and blood vessels are correlated with Alzhemer’s Disease (AD) and can be identied with non-invasive retinal imaging. We show that a denoising diffusion probabilistic model (DDPM) can generate realistic and unique synthetic retinal images. Unimodal convolutional neural networks (CNNs) for predicting Positron Emission Tomography (AmyloidPET) biomarker status were pretrained on synthetic data and finetuned on real data or trained solely on real data. Multimodal classifiers combined unimodal CNN predictions with patient metadata. Our method for generating and leveraging synthetic data has the potential to improve AmyloidPET prediction. Our best unimodal and multimodal classifiers were not pretrained on synthetic data, however pretraining with synthetic data slightly improved classification performance for two out of the four modalities. and integration of metadata in multimodal CNNs achieved best results. Class activation maps show that CNNs for predicting AmyloidPET status can learn features at clinically relevant structures in the retina.