Abstract
Purpose :
To develop a deep learning model for predicting axial lengths (AL) of eyes using optical coherence tomography (OCT) images.
Methods :
We retrospectively included patients with axial length (AL) measurements and OCT images taken within 3 months. We utilized a five-fold cross-validation with the ResNet 152 architecture, incorporating horizontal OCT images, vertical OCT images, and dual-input images. The mean absolute error (MAE), R-squared (R2), and the percentages of eyes within error ranges of ±1.0, ±2.0, and ±3.0 mm were calculated.
Results :
A total of 9,064 eyes of 5,349 patients (total image number of 18,128) were included in the analysis. The average AL of the eyes was 24.35 ± 2.03 (range: 20.53 – 37.07). Using horizontal OCT images, deep learning models predicted AL with MAE of and R2 of 0.704 and 0.784, respectively. Using vertical OCT images, deep learning models predicted AL with MAE of and R2 of 0.698 and 0.785, respectively. Using both horizontal and vertical OCT images, dual-input models predicted AL with MAE of R2 of 0.662 and 0.809, respectively. The dual-input models showed 89.33 %, 98.49 %, and 99.56 % accuracy in error margins of ±1.0, ±2.0, and ±3.0 mm.
Conclusions :
A deep learning-based model accurately predicts AL from OCT images. The dual-input model showed the best performance, demonstrating the potential of macular OCT images in AL prediction.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.