Abstract
Recent application of deep learning in medical image achieves expert-level accuracy. However, the accuracy often degrades greatly on unseen data, for example data from different device designs and population distributions. In this work, we consider a realistic problem of domain generalization in fundus image analysis: when a model is trained on a certain domain but tested on unseen domains. Here, the known domain data is taken from a single fundus camera manufacture, i.e. Canon. The unseen data are the image from different demographic population and with distinct photography styles. Specifically, the unseen images are taken from Topcom, Syseye and Crystalvue cameras. The model performance is evaluated by two objectives: age regression and diabetic retinopathy (DR) classification. We found that the model performance on unseen domain could decrease significantly. For example, the mean absolute error (MAE) of age prediction could increase by 57.7 %. To remedy this problem, we introduce an easy-to-use method, named enhanced domain transformation (EDT), to improve the performance on both seen and unseen data. The goal of EDT is to achieve domain adaptation without using labeling and training on unseen images. We evaluate our method comprehensively on seen and unseen data sets considering the factors of demographic distribution, image style and prediction task. All the results demonstrate that EDT improves the performance on seen and unseen data in the tasks of age prediction and DR classification. Equipped with EDT, the \(\mathrm{R}^2\) (coefficient of determination) of age prediction could be greatly improved from 0.599 to 0.765 (n = 29,577) on Crystalvue images, and AUC (area under curve) of DR classification increases from 0.875 to 0.922 (n = 1,015).
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call