Abstract
PURPOSE: To develop a deep learning (DL) model for segmenting retinal hard exudates (HE) from optical coherence tomography (OCT) scans. METHODS: A modified U-Net architecture was trained on manually segmented OCT B-scans of retinal HE. The training set included 1,811 OCT scans from 15 patients with diabetic retinopathy or branch retinal vein occlusion. The model was evaluated using Dice coefficient and accuracy in idependant test set, and its HE area and volume predictions were compared to manually measured HE areas from a previous clinical study. Additionally, a 2D projected image was generated from the 3D structure of the predicted HE. RESULTS: The DL model achieved a Dice coefficient of 0.721 and an accuracy of 99.9% on the test set. There was a moderate correlation between model-predicted HE volume and area and manually measured HE area from fundus photographs (R = 0.589 and 0.618, respectively; both P < 0.001). The projected 2D image generated from the model accurately visualized HE details, demonstrating better structural information compared to fundus photographs. CONCLUSION: The proposed DL model enables accurate segmentation of retinal HE, offering volumetric data with both horizontal and vertical structural information, which enhances visualization and quantification compared to traditional 2D imaging.