Optimization of Deep Neural Network Model for Embedded Applications: A Case Study of Retinal Vessels Segmentation

Medical science is an area where deep neural networks (DNNs) have been successfully applied for disease diagnosis based on image analysis. The accurate diagnosis of different eye diseases requires retinal vessels segmentation in retinal fundus image. The focus of this work is the improvement of innovative, lightweight deep learning (DL) models that can be used on embedded systems (ESs) for eye diseases diagnosis in the early stage at the point of care in clinics and general hospitals. This work aims to optimize the encoder-decoder-based architecture to show the possibility of porting a DL model onto ESs but without compromising on segmentation performance. The performance was evaluated based on open-source retinal image databases: DRIVE, and CHASE_DB1. The proposed model achieved encouraging evaluation metrics in addition to attributing a significantly lower number of parameters. Moreover, the proposed model outperformed the previous related works while its computational complexity is much lower.