UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Explainable diagnosis using deep-learning : classification of fundus images as a promising tool for early detection of Alzheimer's disease Delavari, Parsa


It has recently been shown that Alzheimer’s Disease (AD) can be detected using machine learning methods and based on retinal fundus images, which are readily available in contrast to conventional diagnosis methods, which are highly invasive and expensive. This study aims to design and validate a methodology for explainable diagnosis of AD based on fundus images using deep learning. Because of the rarity of positive AD labels in the available fundus datasets, patient’s sex was alternatively used as a case study since this trait is also invisible to the experts while convolutional neural networks (CNNs) can predict it. This thesis proposes a novel three-phased methodology to investigate the features in retinal fundus images that enable the CNNs to classify sex. In the first phase, a CNN model is trained and tested on the sex classification task using ODIR and DOVS datasets with 3,146 and 1,600 normal images, respectively. In the next phase, deep learning interpretation techniques were used to hypothesize possible sex differences in the retina. In the third and last phase, the hypotheses were tested using an independent dataset called VCH comprising 500 healthy fundus images. The model achieved AUC scores of 0.68 (95% CI: [0.63, 0.72]) and 0.78 (95% CI: [0.73, 0.84]) on ODIR and DOVS test sets respectively. CNN interpretation results demonstrated that the optic disc, retinal vessels, and fovea are the main anatomical parts attended by the model to predict sex. Feature visualization results followed by statistical tests showed that males have more prominent retinal vasculature by showing more nodes (p=.014), more branches (p=.016), and a greater total length of branches (p=.020) in the retinal vessel network as well as a higher superior temporal vessel coverage (p=.027). Also, the peripapillary area was darker in males compared to females (p=.008). The results confirm the capability of the proposed methodology in identifying the retinal features that are relevant to the task on which the CNN model is trained. The same approach can be extended to other clinically important areas, such as AD diagnosis, considering the ability of CNNs to detect this label.

Item Media

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International