Acceleration of convolutional neural networkbased diabetic retinopathy diagnosis system on fieldprogrammable gate array

Abstract

Diabetic  retinopathy  (DR)  is  one  of  the  most  common  causes  of blindness. The  necessity  for  a  robust  and  automated  DR  screening  system  for  regular examination  has  long  been  recognized  in  order  to  identify  DR  at  an  early stage.   In   this   paper,   an   embedded   DR   diagnosis   system   based   on convolutional  neural  networks  (CNNs)has  been  proposed  to  assess  the proper stage of DR. We coupled the power of CNN with transfer learning to design our model based on state-of-the-art architecture. We preprocessed the input data, which is color fundus photography, to reduce undesirable noise in the image. After training many models on the dataset, we chose the adopted ResNet50  because  it  produced  the  best  results,  with  a  92.90%  accuracy. Extensive experiments and comparisons with other research work show that the  proposed  method  is  effective.  Furthermore,  the  CNN  model  has  been implemented  on  an  embedded  target  to  be  a  part  of  a  medical  instrument diagnostic  system.  We  have  accelerated  our  model  inference  on  a  field programmable   gate   array   (FPGA)   using   Xilinx   tools.   Results   have confirmed  that  a  customized  FPGA  system  on  chip  (SoC)  with  hardware accelerators  is  a  promising  target  for  our  DR  detection  model  with  high performance and low power consumption.