Fourier Ptychographic Reconstruction Method of Self-Training Physical Model

Round 1

Reviewer 1 Report

Good luck.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer:

I appreciate your proposed revisions, I will display the modification result below, please review it.

 

Point 1: What is the meaning of the phrase “Fourier ptychographic…” in the title of the work? I understand there is a reference to [1], but is there a non-Fourier ptychography? Indeed, in practice, the diffraction integral most suitable for a particular experiment is used, which expresses the field on the detector in terms of the field distribution on the surface of the object. As a rule, in one form or another, the calculation of the diffraction integral is reduced to the calculation of direct and inverse Fourier transforms.

Response 1: Fourier ptychography FPM is a computational imaging method that mainly solves the contradiction between large field of view and low resolution in optical imaging systems, it refers to: A series of low-resolution intensity images collected by the Fourier ptychographic microscopy system are combined with the reconstruction algorithm to reconstruct high-resolution intensity and phase images from the collected low-resolution images. The imaging process is: The samples are illuminated in turn at different LED incident angles, and a set of low-resolution ( LR ) images of different spatial spectra are captured accordingly. Then, these spectral sub-apertures obtained from the low-resolution images are spliced together in the Fourier domain to reconstruct the entire high-resolution (HR) spatial spectrum. In short, only the low-resolution intensity image corresponding to the high-resolution spatial spectrum sub-aperture image is obtained, and the low-resolution intensity image is used to restore the high-resolution spatial spectrum. I've highlighted ‘Introduction’ of the article that covers this section.

 

Point 2: What new things did the authors add in this work to what was already published earlier - https://doi.org/10.3390/s22031237 (reference [51])?

Response 2: The network structure in reference [51] is three-part and all are reconstructed in the framework of CNN; in this paper, a new reconstruction structure, Transformer network structure, is adopted. The network structure adopts a modular design, which is easier to capture global information than CNN. It is worth mentioning that the method in this paper is not an improvement based on reference [51], but independent of this method.

 

Point 3: Can the SwinIR network be extended to similar areas of quantitative phase imaging, such as electron microscopy (https://doi.org/10.3390/photonics9110815)?

Response 3: We haven't delved into this part, but we're extending it to the field of image reconstruction for now.

 

Point 4: In addition to the mentioned algorithm by R. Gerchberg and W. Saxton (for some reason abbreviated to the letters G-S), there is also a large class of methods based on solving the intensity transfer equation or similar to it. I would like the authors to mention/comment on them in the Introduction as well (https://doi.org/10.1016/j.optlaseng.2020.106187; https://doi.org/10.3390/s22051765; https://doi.org/10.1038 /s41566-021-00760-8).

Response 4: Thank you for your advice. I have added this part to the introduction.

 

Point 5: What does a high recovery rate mean? Often, modern methods based on ptychography lose out to other methods of quantitative phase microscopy in terms of object recovery time (tens of minutes).

Response 5: Sorry, I didn't mention the high collection rate in the article. I don 't know what the teacher you mean? If it is PSNR (Peak Signal to Noise Ratio), the value range of PSNR is 0-100. The larger the value of PSNR, the smaller the distortion and the better the image quality. If it is the learning rate, then the learning rate is a preset value, which is a value that has been debugged before the experiment and is suitable for the network.

 

Point 6: Please reconsider Figure 3 (upper part), the path of rays through the MO is not very clear

Response 6: Thank you for your advice. I have modified in text.

Author Response File: Author Response.pdf

Reviewer 2 Report

Manuscript ID: applsci-2231309
Title:   Fourier ptychographic reconstruction method of self-training 2 physical model

Author(s):  Xiaoli Wang, Yan Piao, Yuanshang Jin, Jie Li, Zechuan Lin, Jie Cui, and Tingfa Xu

 

The objective of this article is to propose a Fourier picographic reconstruction method based on a physical model of self-training. The authors introduce the SwinIR network into the field of super resolution as the main Fourier stack microscopic reconstruction network where they customize the training process and modify their input into a two-channel input.

  Processing is done by the self-training physical model's stack Fourier microscopic imaging (STPM-FPM) algorithm, and the data set is set to train the network. The results are mainly shown on amplitude and phase images with and without noise. But some details are not clearly established.

I would recommend making revisions to the manuscript and addressing the comments and some questions to be answered listed below:

 

Given the scope of Applied Sciences Journal, this paper needs to properly define all acronyms and terms throughout the article.

Define all variables appropriately in the equations.

The authors state that the self-training SwinIR network structure is superior to the traditional algorithm in terms of reconstruction quality and speed. However, in the images of figures 17 and 18 for the proposed STPM-FPM method, it is qualitatively observed that details are lost or are seen as blurred images. Please explain this issue.

What is the cost/benefit concerning with  reconstruction algoritm proposed in this work.?

 The authors claim:STPM-FPM that uses fewer low-resolution images to reconstruct them with high resolution intensity and phase. Clarify how they improve the illumination angle, the sampling rate, the reduce of artifacts?

Include the area used in the field of view, and the depth of focus value.

Point out if at any point in the reconstruction of the image the wavefront is corrected and in which equations it is introduced?

Indicate the information of the synthetic numerical aperture(s) that was used in the algorithms.

Please add the performance obtained with SwinIR proposed in this work applied to FPM in terms of dB.

The accuracy and error of the amplitude and phase images needs to be included.

Author Response

Dear Reviewer:

I appreciate your proposed revisions, I will display the modification result below, please review it.

 

Point 1: Given the scope of Applied Sciences Journal, this paper needs to properly define all acronyms and terms throughout the article.

Response 1: Thank you for your advice. I have modified at Introudction, Section 2.1.

 

Point 2: Define all variables appropriately in the equations.

Response 2: Thank you for your advice. I have revised at Section 2.1

 

Point 3: The authors state that the self-training SwinIR network structure is superior to the traditional algorithm in terms of reconstruction quality and speed. However, in the images of figures 17 and 18 for the proposed STPM-FPM method, it is qualitatively observed that details are lost or are seen as blurred images. Please explain this issue.

Response 3:  For Figure 17 and 18, the method used in this paper is mainly to deal with the interference condition-noise. It can be clearly seen from the results that Figure 17 and 18 contain much less artifacts than the previous methods.

 

Point 4: What is the cost/benefit concerning with reconstruction algorithm proposed in this work.?

Response 4: The method used in this paper is the FPM reconstruction method of self-training physical model, which effectively solves the decisive factors of different network models for reconstruction quality. At the same time, the idea of transfer learning and deep learning can fully improve the quality of reconstructed images. Compared with other traditional and neural network-based reconstruction algorithms, this paper has a more obvious denoising effect on noise. However, the deep learning reconstruction method is affected by its own reconstruction network, such as the adjustment of hyperparameters, the selection of loss function, the structure of the network and other factors. Therefore, how to improve the stability of the deep learning reconstruction method remains to be studied. At the same time, the cycle time of training based on deep learning method is long, and how to lighten the model is one of the key research directions in the future.

 

Point 5: The authors claim:STPM-FPM that uses fewer low-resolution images to reconstruct them with high resolution intensity and phase. Clarify how they improve the illumination angle, the sampling rate, the reduce of artifacts? Include the area used in the field of view, and the depth of focus value.

Response 5:  In this paper, the idea of transfer learning is adopted, and the SwinIR network is introduced. In terms of image use, it is based on the data set of simulation and the data set of real collected images. The objective lens of the FPM imaging system with a numerical aperture NA of 0.13 is used to collect images in the experiment. The light intensity image is recorded by a 2560 × 2560 pixel (6.5 um pixel size) scientific CMOS camera. A planar array is used as a 13 × 13 programmable light source element LED, and the illumination wavelength is 505 nm at 100 mm below the sample to provide illumination. I've highlighted 4.1.3 of the article that covers this section.

 

Point 6: Point out if at any point in the reconstruction of the image the wavefront is corrected and in which equations it is introduced?

Response 6: What has a big impact on FPM systems is position correction, which was done when the analog image database was made.

 

Point 7: Indicate the information of the synthetic numerical aperture(s) that was used in the algorithms.

Response 7:  Numerical aperture (s) is used to build a real image data set. The size of the numerical aperture (s) is: 0.13.

 

Point 8: Please add the performance obtained with SwinIR proposed in this work applied to FPM in terms of dB. The accuracy and error of the amplitude and phase images needs to be included.

Response 8:  Thank you for your advice. I have added ' dB' after the PSNR values of Table2 and Table3.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors took into account all comments. The article may be published.