Separate Detection of Stromal and Epithelial Corneal Edema on Optical Coherence Tomography Using a Deep Learning Pipeline and Transfer Learning

Round 1

Reviewer 1 Report

this is a very interesting work regarding corneal edema grading.

overall few works were done with oct.

few remarks for the authors:

1 it is not clear what are the clinical parameters of corneal edema used for corneal edema grading

2 is there any effect of the oct device resolution?

3 how you took into concederation minor corneal opacitis that can be created because of chronic edema or even in the stroma after surgery?

Author Response

Dear Reviewer, 

We would like to thank you for judging our work. We addressed each of your comments and modified the manuscript accordingly.We believe it has significantly improved the quality of our manuscript.

1 it is not clear what are the clinical parameters of corneal edema used for corneal edema grading

Thank you for this comment. Overall, in an attempt to create a precise deep learning model, we tried to limit the subjectivity of our data labelling and therefore not to rely on clinical evaluation. Regarding the development set, we assumed patients who underwent DMEK surgery the day before had total corneal or near-total edema which is usually the case. Regarding the validation set, as we used images from patients having minimal edema, we used the difference between post-operative and preoperative central corneal thickness (DCCT) as ground truth quantification of edema. This is the most objective parameter available as other clinical classification are usually insufficient to characterize mild corneal edema. We are aware of the limitations of this approach. First, only central edema is quantified, and localized peripheral edema might be overseen. Second, in FECD, the Descemet membrane is usually thickened and can falsely increase the DCCT value. Nonetheless, we believe all deep learning models should be developed using the most objective data to avoid mimicking our clinical flaws. It would however be interesting to compare our results to clinical grading and this will be the subject of a future work. Precisions on all those matters have been included in the text.

2 is there any effect of the oct device resolution?

Thank you for this question. Currently this model has been developed only for the Avanti Optovue device. It is certainly less accurate on other devices as it was not trained with other types of images. If trained with other kind of images, we suspect lower resolution images to yield less precision in cases of very mild edema but similar results in cases of more advanced edema. We are currently working on an extension of this model to other devices, and it will be the subject of a future work. We added comments on that matter in the discussion section.

3 how you took into concederation minor corneal opacitis that can be created because of chronic edema or even in the stroma after surgery?

That is a great question. Currently, all minor opacities visible in the development set images were labeled as edema as all other corneal pixels. Interestingly, when minor opacities exist in non-edematous corneas the model predicts no edema in these locations. This is due to the fact that the model is confident that there is no edema in the surrounding pixels and therefore infer the “normal” class to these pixels. We agree that in future versions of the model, opacities should be labeled separately. We added a comment in the discussion about corneal opacities.

Reviewer 2 Report

The author used a deep learning pipeline for separating epithelial edema and stromal edema via OCT images of the cornea to classify normal, minimal edema, and important edema. The author also adopted the transfer learning to minize the training images. This work was rational because the gray levels of corneal epithelium and those of corneal stroma were different. However, there are some concerns for the author to address.

1. Whole corneal thickness can be simultaneously determined by anterior segment OCT and Scheimpflug tomography, and the former can also determine the corneal epithelial thickness. The clinical value seems not high via this effort of the author.
2. The developed model just distinguishes normal, mild edema, and significant edema rather than provides an exact value of corneal thickness and the location of corneal edema.
3. Figures 6-8 only show a sagittal view or cross section of the cornea. Do the author adopt whole cornea pixels (3D) or just only pixels of one sagittal plane of cornea (2D) for determining the corneal edema?
4. There are many instruments useful to determine central, paracentral, and peripheral corneal thickness, such as specular microscopy, Scheimpflug tomography, and anterior segment OCT. However, in this study, there is lack of whole corneal thickness to validate or correlate the result of this study. Could the author obtain more points with known local corneal thickness from central, paracentral, and peripheral cornea  for a better proof of identificating mild and local corneal edema?

Author Response

Dear Reviewer, 

We would like to thank you for judging our work. We addressed each of your concerns and modified the manuscript accordingly. We believe it has significantly improved the quality of our manuscript.

The author used a deep learning pipeline for separating epithelial edema and stromal edema via OCT images of the cornea to classify normal, minimal edema, and important edema. The author also adopted the transfer learning to minize the training images. This work was rational because the gray levels of corneal epithelium and those of corneal stroma were different. However, there are some concerns for the author to address.

1. Whole corneal thickness can be simultaneously determined by anterior segment OCT and Scheimpflug tomography, and the former can also determine the corneal epithelial thickness. The clinical value seems not high via this effort of the author.

Thank you for this comment. Few studies have been published on the detection of corneal edema on OCT images only, and the best methodology to do so is yet to be determined. Therefore, in this work we describe a way to improve the deep learning methodology for that purpose. Our results show the performance of a model based only on OCT images. We agree it would be interesting to combine whole corneal thickness and epithelial thickness mapping measurements to the model to further increase its performance and this could be the subject of a future work. Precisions on that matter have been included in the introduction and discussion sections.

1. The developed model just distinguishes normal, mild edema, and significant edema rather than provides an exact value of corneal thickness and the location of corneal edema.

Thank you for this comment. In fact, our model does not provide a class diagnosis of “normal”, “mild” or “significant” edema. It predicts the presence of edema at the pixel level each sagittal image. This information can then be derived using post-processing to get an idea of the location and quantity of edema. Indeed, the location of edema can be inferred from the heatmap prediction on each radial scan. By combining the results of all radial scans, we could provide an en-face mapping of edema on the whole cornea. Even though the model is not exactly designed to quantify corneal edema but rather to detect it, one can get an idea of the severity of edema through post-processing using the number of pixels detected as edema compared to the number of total corneal pixels. Thus, our model is a tool to detect edema on images. We agree that the optimal post-processing tools to provide the most clinically valuable results need to be determined. We added precision on that matter in the discussion section.

1. Figures 6-8 only show a sagittal view or cross section of the cornea. Do the author adopt whole cornea pixels (3D) or just only pixels of one sagittal plane of cornea (2D) for determining the corneal edema?

Thank you for this interesting question. Currently, the model only uses 2D images to make the predictions. The scanning protocol used, only provides 8 radial scans separated by 22.5° which does not allow to reconstruct a proper 3D volume. The scanning protocol available to acquire 3D volumes on this device uses a lower resolution and would certainly yield worse results. Nonetheless, we used the model to detect edema on each of the 8 radial scans for each patient and averaged the Edema Fraction over all scans per patient to get an idea of the whole corneal edema. We clarified the method section on the matter.

Future works could use more a complicated model to take into account all scans simultaneously. A comment on that matter was added in the discussion section

1. There are many instruments useful to determine central, paracentral, and peripheral corneal thickness, such as specular microscopy, Scheimpflug tomography, and anterior segment OCT. However, in this study, there is lack of whole corneal thickness to validate or correlate the result of this study. Could the author obtain more points with known local corneal thickness from central, paracentral, and peripheral cornea  for a better proof of identificating mild and local corneal edema?

Thank you for this comment. We agree it would be interesting to compare our results to whole corneal thickness mapping to validate the detected locations of edema. In this study, we focused on a common clinical problem which is the determination of mild edema in FECD patients. FECD edema usually starts in the central region. Thus, we believe the current approach using the average central corneal thickness in the 3 central millimeters is appropriate for that specific problem. Nonetheless, in order to validate the model to other use cases, we will indeed compare the results to full corneal thickness mapping in further studies. We adjusted the discussion in consequence.

Reviewer 3 Report

 The manuscript titled "Separate Detection of Stromal and Epithelial Corneal Edema on Optical Coherence Tomography Using a Deep Learning Pipeline and Transfer Learning" presents an interesting concept on diferential corneal edema detection, but the manuscript shows insufficient scientific content, weak introduction and methodology and not clear conclusions.

I can not recommend this article for publication because of major reasons:

1. Corneal imaging resolution does not provide spatial information enough for differential segmentation.
2. Insufficient trainning set of images and subjects.
3. Poor introduction.
4. Poor metholodogy description. Deep learning architecture is not clearly described in order to provide other researchers to replicate it.
5. Image processing is poorly described.
6. Results are poorly described.
7. The manuscript does not present conclusions that support worthy and original research.

Author Response

We would like to thank the reviewer for judging our work. We answered to each of your comments and adjusted the manuscript accordingly. We also provide a supplementary table describe the full model architectures. We believe it has significantly improved the quality of our manuscript.

1. Corneal imaging resolution does not provide spatial information enough for differential segmentation.

Thank you for this comment. Despite an axial resolution of 5 μm, it is possible that very minimal edema might not be detectable on corneal OCT. However, mild corneal edema induces specific features visible on OCT images by trained experts. Indeed, we can see epithelial and stromal hyperreflectivity, disorganized collagen lamellae, corneal folds and shadows. As it is visible by the human eye, a deep learning model should be able to perform at least as well in the detection of those specific features. In fact, our previous published work on the subject confirmed that hypothesis as it allowed to detect corneal edema in different clinical settings with a good correlation with the clinical diagnosis. We added the device resolution in the method section and discussed the limitations induced by image resolution to the discussion section.

2. Insufficient trainning set of images and subjects.

Thank you for this comment. We agree that the performance of the model could be improved by adding training data. However, we wanted to build on our previous published work and show how a different methodology could improve the results using the same exact training set. In that matter, we did not increase the training set size. Another goal of this work was to show that transfer learning can allow to build useful specific tools with very limited training sets. Therefore, we deliberately used a small fraction of our original training set to train the model for epithelium segmentation. Even though the precision of epithelial segmentation could be improved with a larger training set, the current model provided results sufficient for the required task. The purpose of this paper was more to describe a novel methodology rather than a final model directly usable in clinical practice. To get there, we would indeed need to increase the training set size, and this will be the subject of a future work. Validation on a larger cohort will also be tested in a future work. We added clarifications on these matters in the discussion sections

3. Poor introduction.

Thank you for that comment. We added details in the introduction regarding OCT and other imaging techniques available to detect corneal edema with the relevant references.

4. Poor metholodogy description. Deep learning architecture is not clearly described in order to provide other researchers to replicate it.

We thank the reviewer for that comment. We added the full architecture details in a supplement table. The rest of the methodology including the training hyperparameters and data augmentation process is fully described in the manuscript.

5. Image processing is poorly described.

Thank you for this comment. The image processing section of the manuscript describes the only process applied to our images which was a lateral cropping of larger images to make all images the same size. We deliberately did not use any preprocessing technique on the development set to make the model more robust to signal variation, noise and artifacts encountered in OCT images in routine clinical setting. Therefore, no other pre-processing or post-processing was necessary on the validation images as they were comparable to the development set. We added sentences stating that no other processing technique has been applied.

6. Results are poorly described.

Thank you for this comment. We described the full training results details with the learning curves of each model. Also, the main outcome results (Edema Fraction) are described for each group of the validation set and comparison through statistical testing is also provided. Several examples of the color heatmap predicted by the previous and current model allow the reader to visualize the results in cases of minimal edema with their corresponding DCCT provided in the figure caption. We added descriptive details of each visual example.

7. The manuscript does not present conclusions that support worthy and original research.

Thank you for that comment. We added a conclusion section explaining how our work provides an improvement of the deep learning methodology to detect corneal edema and highlights the problem of bias induced by the combined analysis of epithelium and stroma. This finding might be of interest for other researchers working on the subject.

Round 2

Reviewer 2 Report

The author has replied to all my concerns.

Author Response

Thank you

Reviewer 3 Report

Authors have improved their manuscript in the sense of my comments,but the revised introduction is still poor in references.

Discussion section does not compare their study with previous work.

Author Response

Dear Reviewer, 

Thank you for theses additional comments. As suggested, we added 5 new relevant references to the introduction and believe it has improved its quality. We also compared our results to those of two recent studies tackling the problem of corneal edema detection.