Segmentation of Retinal Layers for Detecting Accumulated Fluid Regions using a U-Net Mx-Net Architecture

In early 1990's the potential of OCT technology was recognized. Segmenting retinal layers, to detect fluid and identifying eye diseases is a challenging task. There exist several automated techniques like image processing, machine learning. Unfortunately, these often suffer from error as well as inefficiency in computation and require further interpretation of human expertise. In this work, a novel network architecture has been proposed using U-Net with Apache Mx-Net framework for segmenting the fluid region inside the retina. The main advantages of this approach are it can be performed in real-time, reduce training time and provide speedy identification of diseases. It removes the trade-off between patch size and localization accuracy. We added ResConv and ResDeconv block with the traditional U-Net architecture which removed the gradient vanishing problem and Mx-Net enabled running the whole program within a very short period with fast tuning. We have created a hybrid forward method to stack the layers. As a scalable approach, this model can be easily put into any adaptive learning model or a bigger CNN and pass the model to convert data function. The best part about feature visualization is that it kind of tells us how our eye detects features. The experiment outperforms on 10 patients which signifies that our model is efficient enough for retinal boundary segmentation and fluid detection. The model obtained a testing accuracy of 89.8%.