A CNN-Based Framework for Automatic Vitreous Segemntation from OCT Images

Accurate segmentation of the vitreous region of retinal images is an essential step in any computer-aided diagnosis system for severity grading of vitreous inflammation. In this paper, we developed a framework to automatically segment the vitreous region from optical coherence tomography (OCT) images of uveitis eyes using fully convolutional neural network (CNN), U-net model. The CNN model consists of a contracting path to capture context and an expanding path for precise localization and utilizes the binary cross entropy (BCE) loss. The model has been tested on 200 OCT scans of eyes having different grades of uveitis severity (0–4). The developed CNN model demonstrated not only high accuracy of vitreous segmentation, documented by two evaluation metrics (Dice coefficient (DC) and Hausdorff distance (HD) are 0.94 ± 0.13 and 0.036 mm ± 0.086 mm, respectively), but also requires a small number of images for training. In addition, the training process of the model converges in few iterations, affording fast speed contrary to what is expected in such cases of deep learning problems. These preliminary results show the promise of the proposed CNN for accurate segmentation of the vitreous region from retinal OCT images.