Segmentation of Retinal OCT B-scans Using a Variable Kernel U-Net Architecture for Diagnosis of Diabetic Macular Edema

Automated segmentation of anatomical and pathological features in retinal optical coherence tomography (OCT) B-scans is critical for ocular disease diagnosis. In this paper, we present a Variable Kernel U-Net (VKU-Net) architecture for joint retinal layer segmentation and fluid detection in OCT B-scans with diabetic macular edema (DME). The proposed method is trained and validated on a publicly available OCT dataset acquired from 10 patients with DME. The Dice Coefficient and the Precision-Recall scores are used to assess the network performance. Compared to U-Net with fixed filter size, VKU-Net achieves comparable performance utilising less than half of the parameters. Furthermore, the consistency of these quantitative results with the automated segmentation maps is investigated in order to understand the scope of our proposed method.