LiteVessel: In-Depth Exploration of Lightweight Deep Neural Network Models for Retinal Vessel Segmentation

Abstract

Deep learning has been used over the past decade for diagnosis applications in healthcare including ophthalmology. The integration of deep learning models with embedded systems to attain real-time processing of diagnosis becomes ineffective due to the resource constraints of embedded systems and higher computation and memory requirements of DNNs. To overcome this issue, this work aims to optimize an encoder–decoder architecture to demonstrate the potential for porting a DL model to any general embedded platform for eye disease diagnosis in the early stage. In this paper, we tested different model architectures to reduce the computation complexity of the DL model without compromising performance metrics. To train and test our optimized models, we utilized available databases of retinal images such as DRIVE, CHASE_DB1, and STARE. Although the computational complexity was much lower, the developed models achieved competitive performance compared with the existing state-of-the-art. Furthermore, we implemented a cross-training approach, and the findings illustrate the generalizability and resilience of the methods presented. The reduced number of parameters, computational complexity, and enhanced segmentation performance of retinal vessel segmentation make the proposed methods suitable for use in automated diagnostic systems.