Predicting axial length and refractive error by leveraging focus settings from widefield fundus images

Abstract

Purpose : Recently deep learning (DL) has been used for predicting more than retinal information from fundus images: for example, age, blood pressure, etc. In this pilot study, we developed an algorithm to predict the axial length (AL) and spherical equivalent (SE) refractive error of the eye from an image of the fundus by using transfer training from the focus settings (FS) used in the fundus cameras.

Methods : This retrospective study used 5291 widefield true color images acquired using CLARUS™ 500 and 700 (ZEISS, Dublin, CA) along with the FS in diopter (D) recorded during the acquisition. FS of the camera used in the capture of the images was related to the refractive error and AL is regarded as one of the determinants of refractive error [Mutti et al. IOVS 2007; 48(6)]; hence, the model trained using FS was used to transfer train AL and SE prediction. This algorithm was trained in two steps: 1) training a DL model to predict FS, and 2) transfer learning to predict AL and SE.
A 3-channel ResNet-50 with softmax activation pre-trained on ImageNet dataset is trained on the fundus to predict FS with 80:20 split between training and testing. For transfer learning to predict AL and SE, 162 images along with AL measured on IOLMaster® 700 (ZEISS, Jena, Germany) and 367 images along with SE assessed using VISUREF® (ZEISS, Jena, Germany) were used. The following test sets were used for evaluating the performance: 1) 41 images with corresponding AL; 2) 94 fundus images and corresponding SE. The mean absolute error (MAE) was used to measure the performance.

Results : The model trained to predict the FS has a MAE of 0.49±0.27D. The transfer trained algorithm has a MAE of 0.61±0.41mm and 0.73±0.56D for predicting AL and SE, respectively. Figure 1 shows some of the examples of predictions from the proposed algorithm. The model accuracies for different error margins are shown in Table 1.

Conclusions : In this pilot study, transfer learning from focus setting enabled prediction of axial length and refractive error. To our knowledge, the feasibility of using DL to predict axial length from retinal images has not been previously demonstrated. In the future, these DL methods may provide solutions for customizing the measurements in fundus cameras for an individual eye by estimating AL instead of using an optical model.

This is a 2020 Imaging in the Eye Conference abstract.

 

Figure 1. Results from the proposed algorithm

 

Table 1. Summary of results for different error margins