1 Reviewer 1
   2 ----------
   3 This paper introduces an FPGA acceleration solution of cell imaging and classification for quantitative phase asymmetric-detection time-stretch optical microscopy. With the help of quantitative phase imaging, the system not only improves the classification accuracy by 2% to 4%, but also enables real-time processing with a relatively high throughput. This could help the related research areas taking advantages of this classification scheme.
   4 Comments:
   5 1.	According to the paper, the authors do well in transverse comparison (CPU, GPU), but relatively little attention has been paid to the horizontal comparative research, including the declaration of the related work and contrast.
   6 2.	The proposed FPGA acceleration system mostly adopts the general module or method, such as frequency domain module and SVM classification module, etc. it will be better to have some innovation on this point.
   7 
   8 Reviewer 2
   9 ----------
  10 This paper presents an FPGA-based accelerator for
  11 cell image classification. The contributions are
  12 solid and the presentations are clear and
  13 concise. I recommend the acceptance of the paper.
  14 
  15 One minor suggestions for the paper presentation
  16 is to add a paragraph/section on host/FPGA communication
  17 bandwidth. There are multiple places in the paper stating
  18 that eventually that system throughput would be limited
  19 by PCIe bandwidth, and Equation (15) even give formula
  20 for that. However, I could never see any real data
  21 in terms of the bandwidth actually consumed (in GB/s,
  22 for 1/2/3/4 FPGAs). I believe it is an important data point.
  23 
  24 Reviewer 3
  25 ----------
  26 This paper presents a novel method to accelerate cell classification. The work is very detailed in terms of both the resulting architecture, as well as its performance and power characteristics. Overall this seems like a very promising architecture. The paper is well-written, and suitable for an FPT audience.
  27 
  28 Reviewer 4
  29 ----------
  30 This is a well-written paper that implements Quantitative Phase Imaging (QPI) on FPGAs to support high throughput data analysis for quantitative asymmetric-detection time-stretch optical microscopy (ATOM).
  31 The pixel streams produced by ATOM are first processed by a spatial domain module, which is composed of a background subtraction kernel, an intensity normalization kernel, and a complex phase shift extraction kernel. The results of the spatial domain module are forwarded to a frequency domain module, which reduces high frequency noises by combining a forward 2D FFT, a low pass filter, and an inverse 2D FFT. Finally, 
  32     QPI processed images are classified using Support Vector Machines. The architecture is implemented using Maxeler tools and achieves a speed-up of 9.4x over a 6-core Intel Xeon CPU, and a speed-up of 3.47x over an Nvidia Tesla K40C GPU when using a single FPGA card. The authors also provide a theoretical scaling of the performance when using multiple FPGA cards.