Abstract
Purpose :
It is well-known that age and other key attributes/risk factors can be predicted from fundus images with DL models specifically trained to do so. The current dogma is that these models need to learn from vast amounts of data to recognise subtle patterns. Instead of training on fundus images, we used a DL model trained only on natural images (pets, cars, food, etc.) and UMAP to investigate the key axes of “perceptual” variation.
Methods :
We used a ResNet10t DL model pre-trained on a dataset of natural images (ImageNet) which was never trained on retinal images. Left eye images of 37,588 White British participants in UK Biobank were passed through the model except for the final classification layer. This yields representations of each image consisting of 512 numbers which capture high-level perceptual features useful for classification of natural images. We can interpret these numbers as a point in 512-dimensional space. Those were then reduced to two dimensions with UMAP which preserves the global structure of the data. (Fig 1) At no point was age considered in this process. Finally, we plotted the UMAP representations and coloured each point by age.
Results :
In the scatterplot (Fig 2), images of participants with similar age are close to each other and there is a noticeable young-old gradient from the bottom left to top right. We verified this impression quantitively by evaluating (12-fold crossvalidation) how well age is predicted from the 2d representations. Using the average age of training points within a radius of 1/3, age of held-out points was predicted with an R2 of 0.1717(±0.0158). A simple linear regression achieved an R2 of 0.1069(±0.0119).
Conclusions :
The high-level perceptual representations captured salient information about age without the DL model being trained to recognise subtle patterns specific to fundus images. That this information was retained by UMAP suggests that age is a key axis of perceptual variation in fundus images. Thus, at least some age-related patterns in fundus images are perceptually obvious rather than subtle. Our work sheds new light on age prediction and DL in fundus imaging generally. The presence of informative, high-level visual patterns might provide a new avenue for interpretable DL models that are less opaque to clinicians. Future work should identify these patterns and investigate other risk factors.
This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.