Authors

Mohsin Challoob, Yongsheng Gao, Andrew Busch, Weichuan Zhang

Abstract

Robust delineation of retinal microvasculature in optical coherence tomography angiography (OCTA) images remains a challenging task, particularly in handling the weak continuity of vessels, low visibility of capillaries, and significant noise interferences. This paper introduces a modulatory elongated model to overcome these difficulties by exploiting the facilitatory and inhibitory interactions exhibited by the contextual influences for neurons in the primary visual cortex. We construct the receptive field of the neurons by an elongated representation, which encodes the underlying profile of vasculature structures, elongated-like patterns, in an anisotropic neighborhood. An annular function is formed to capture the contextual influences presented in the surrounding region outside the neuron support and provide an automatic tuning of contextual information. The proposed modulatory method incorporates the elongated responses with the contextual influences to produce spatial coherent responses for delineating microvasculature features more distinctively from their background regions. Experimental evaluation on clinical retinal OCTA images shows the effectiveness of the proposed model in attaining a promising performance, outperforming the state-of-the-art vessel delineation methods.

Link to paper

DOI: https://doi.org/10.1007/978-3-031-43990-2_67

SharedIt: https://rdcu.be/dnwMr

Link to the code repository

N/A

Link to the dataset(s)

N/A

Reviews

Review #3

Please describe the contribution of the paper

A novel modulatory elongated model that requires no annotated data for training for segmentation of retinal microvasculature in OCTA images. The evaluations shows better performance compared to the non-learning SOTA baselines.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1. A novel modulatory elongated model based on neurophysiological evidences for the primary visual cortex for segmentation of retinal microvasculature in OCTA images.
2. A non-learning approach, which does not require any annotated data for training.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1. The complete algorithm should be presented, not only the modulatory elongated model.
2. Ablation study is missing.
3. Lack of comparison with the learning-based SOTA.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The work might not be reproducible. ROSE-1 is a public dataset and the modulatory function is well-defined, but the complete algorithm for turning a OCTA image into a vessel segmented image is not specified.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
1. In the introduction, the authors should make clear of the type of model proposed. Is it a learning approach? If not, the word the “unsupervised” should be avoided, because it implies “unsupervised learning”. Unsupervised is not the same as non-learning.
2. Should contribution 1 and 3 be combined?
3. What is the complete algorithm that embeds the proposed model for extracting elongated structures? Any pre- or post-processing required to produce the results shown in Fig.1? Is M(x,y) computed at every pixel location?
4. Are the baselines only non-learning approaches? Do they really represent the SOTA for vessel delineation, as mentioned in the introduction? How about supervised learning approaches using available annotated data? Do they do better than the proposed model?
5. Should a metric measuring connectiveness/continuity be included to address the first challenge on the issue of disjoint detection?
6. Ablation study missing. Should contextual influence H and contrast influence S be evaluated separately?
7. Is the designed model only effective for OCTA images? How about ultra-widefield FA or FP images?
8. Is the model scale-invariant for images of different resolutions?
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

5
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

I recommend this paper based on the fact that it is not another deep learning paper. However, the recommendation is weak because evaluation is insufficient, and the complete algorithm for producing Fig1 is missing.
Reviewer confidence

Very confident
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #1

Please describe the contribution of the paper

This paper presents a modulatory elongated model that exploits contextual modulatory effects presented for tuning the responses of neurons in V1. The proposed method in-corporates the elongated responses, neuron responses, at a certain location with either a facilitatory or inhibitory process, contextual information, to reliably enable the segmentation of vasculatures in OCTA images. In general, it is interesting to see such an geometric modeling based method that has been explored to segment OCT microvasculature, which shows practical values.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1. An unsupervised cortical-inspired model has been established to process retinal OCTA images.
2. Contextual oriented filters are elaborated to segment vessel structures and improve the continuty.
3. The proposed method achieves competitive performance compared with other unsupervised approaches.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
1. The conventional geometrical models could have limitation in segmenting capillary structures, which are not fully validated due to the lack of capillary labels in the public dataset. However, it is still interesting to see how it performs.
2. Although the proposed method is an unsupervised approach and the authors only compared with several geometric methods, it is still interesting to compare also with the most recent unsupervised/supervised learning techniques.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The proposed model is clearly presented and thus it is feasible to reproduce the results.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html
1. Other similar models such as the optimally oriented flux filters might be also interesting to be compared, particularly on segmenting tiny structures.
2. The authors showed that the proposed method is better at processing discontinuous segments and provide complete connections. It would be interesting if the authors can further validate its clinical significance based on the improvement.
3. Since small capillaries play more important role in vessel analysis rather than small vessels, the authors are encouraged to provide specific evaluation on capillary segmentations
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

5
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

It is intersting to see that the orientation selective model has been explored to segment microvascular structures and it shows advantages in solving discontinuity issue.
Reviewer confidence

Very confident
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Review #2

Please describe the contribution of the paper

The presented work introduces a novel kernel-based segmentation method for tubular structures. Motivated by physiological evidence of the functioning of the human visual cortex, it extends previous work on elongated vesselness filters (reference 19). To this end, it includes a modulatory signal based on the local contrast and information outside the kernel’s receptive field. The proposed method is benchmarked on the ROSE-1 dataset of retinal OCTA images. It is quantitative and qualitatively shown to outperform seven other classical computer vision methods.
Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The segmentation of OCTA images is challenging and clinically relevant.
- The method and conducted experiments are clearly presented.
- I found the description of the human visual cortex and its use as inspiration for the proposed method interesting.
Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- In several passages, I found the paper difficult to understand. Especially in the beginning the authors dive into theory about the human visual cortex without clearly establishing the ultimate aim of their work. Later, once the work’s purpose became clear and the writing is supported by maths equation and results figures, I found the paper substantially easier to follow. The authors should concisely state their aims and contributions at the beginning of the paper. Additionally, they could consider adding an introductory figure providing intuition of the proposed method.
- As I understand it, the newly introduced kernel that models contextual information (Equations 4 to 6) raises the threshold for a pixel to be deemed vessel if there is a lot of spurious signal in a ring around the region of interest. In the context of OCTA images, this means that the firing of the kernel is particularly supressed in regions of dense capillary networks. Coincidentally, this matches the labelling characteristics in the ROSE-1 dataset. In the ROSE-1 dataset manual annotators have predominantly outlined major vessels while ignoring the capillaries, even though the latter have been shown to be highly relevant for a range of ophthalmological and cardiac diseases. Additionally, I suggest also including the ground truth labels in Figure 3.
- In light of above’s comment, I suggest that the authors replicate their results on additional datasets. At the moment the utility of the proposed method is only demonstrated on a single dataset. There are several other publicly available OCTA datasets with labels, such as the OCTA-500 dataset (Li, Mingchao, et al. “Ipn-v2 and octa-500: Methodology and dataset for retinal image segmentation.” arXiv preprint arXiv:2012.07261 (2020)) and a smaller dataset by Giarratano et al. (Giarratano, Ylenia, et al. “Automated segmentation of optical coherence tomography angiography images: benchmark data and clinically relevant metrics.” Translational vision science & technology 9.13 (2020): 5-5). Beyond OCTA images, there are also public datasets of color fundus images (Staal, Joes, et al. “Ridge-based vessel segmentation in color images of the retina.” IEEE transactions on medical imaging 23.4 (2004): 501-509).
- While I appreciate the advantages of classical kernel-based segmentation methods, I do not think that deep learning-based methods can be entirely ignored. At the very least the authors should aim to include the quantitative results by Ma et al. (reference 2). Accompanying their dataset, they have provided a public code repository with several neural networks for segmentation.
Please rate the clarity and organization of this paper

Good
Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance

The authors describe their method in extensive detail and use a publicly available dataset, making the research highly reproducible.
Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2023/en/REVIEWER-GUIDELINES.html

I believe the authors should address the two main concerns raised in my comments above. On the one hand, they should aim to replicate their results on additional datasets of tubular structures. On the other hand, they should include at least a few learning-based baselines as comparison in order to faithfully reflect the current state-of-the-art in image segmentation.
Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making

4
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

As outlined in my comments above, I am concerned about overfitting of the method to a single dataset and its labelling characteristics. Furthermore, I think that at least a few learning-based baseline have to be included in any contempory segmentation work, considering their dominance in computer vision. Unfortunately, these issues keep me from scoring the work higher.
Reviewer confidence

Confident but not absolutely certain
[Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed

N/A
[Post rebuttal] Please justify your decision

N/A

Primary Meta-Review

Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper presents a novel modulatory elongated model. The method extends elongated vesselness filters to combine elongated responses at a location with facilitated or inhibited processes and contextual information for unsupervised segmentation of retinal vessels in OCTA images. Three reviewers agreed that the method of this paper is novel, but it has weaknesses including insufficient experiments and unclear exposition. This paper exhibits the following issues, and corrections in their future work (if the authors would like to prepare an extension version) will improve the quailty of this paper.
1. The proposed method is demonstrated only on a single dataset and lacks the validation effect of public datasets. (R1,R2)
2. The presentation of the content of the paper is not clear. The authors did not state their objectives and contributions succinctly. The description of the methods is also not intuitive enough. (R2,R3) 3 . The paper only compares with several geometric methods and lacks comparison with the latest unsupervised learning techniques, as well as does not show the performance of deep learning based methods. (R1,R2,R3)
3. The authors failed to further validate their clinical significance of this method. (R1)

Author Feedback

We appreciate the valuable comments and suggestions on improving the manuscript. We address the concerns raised by the reviewers as follows.

(Response to R1,R2,R3): As suggested, we have now evaluated our method in comparison with the benchmarks on another clinical OCTA dataset: OCTA500 (ILM_OPL, 6mm) dataset. Note that the majority of the OCTA_6mm subjects (69.7%) have various retinal diseases (AMD, DR, CNV, CSC, RVO, and others). The proposed method again obtains the best performance over the benchmarks, as follows: the proposed method (ACC= 0.966; FDR= 0.141; PVR= 0.915), SOGGDD [19] (ACC=0.949; FDR=0.238; PVR=0.857), RUSTICO [10] (ACC=0.941; FDR=0.270; PVR=0.835), PCT [11] (ACC=0.949; FDR=0.226; PVR=0.862), SCIRD [12] (ACC=0.952; FDR=0.203; PVR=0.876), PFT [14] (ACC=0.950; FDR=0.212; PVR=0.870), MBT [15] (ACC=0.957; FDR=0.165; PVR=0.898), and RVR [13] (ACC=0.954; FDR=0.188; PVR=0.884).

(Response to R3): The proposed method is a non-learning approach. We did not use any pre-or post-processing to produce results. In Fig.1, our results and ground truths are simply overlaid on original images for better visualization. The delineation results of our method and comparative benchmarks are segmented by selecting the best threshold that returns the highest average 𝐴𝐶𝐶 on each dataset. M(x,y) is computed at every pixel location (x,y).

(Response to R1,R2,R3): This paper introduces a new non-learning method that advances the delineation technology of retinal microvasculature in OCTA images for addressing the limitations in the existing vessel detection methods. We compare our method with the existing non-learning baselines. The proposed method is a single non-learning operator and is not directly comparable with deep learning models, which use hundreds, or thousands of filters constructed in a hierarchical fashion (consisting of multiple modules and layers), and rely on labour-intensive manual labelling and training process. Also, we follow the same evaluation protocol as used in [12], [13], [14], [15] for performance comparison.

(Response to R2): The existing algorithms have achieved a great progress for vessel enhancement and extraction. However, the following problems remain unsolved and need to be overcome for a reliable vasculature delineation for OCTA images. (i) Some vessels and capillaries suffer from the problem of weak continuity, which leads to a disjoint detection in segmented vascular trees. (ii) Due to the poor SNR of OCTA images, some capillaries are presented with an inadequate contrast, causing difficulty in differentiating them from inhomogeneous background. (iii) OCTA images are associated with high noise level, which significantly interferes with vessel structures and makes their boundary irregular. This paper introduces a new modulatory elongated model to advance the delineation methodology of retinal microvasculature in OCTA images via addressing the above problems. The contributions of this work are summarized as follows. (1) A modulatory function is proposed to include two simultaneous facilitatory and inhibitory delineation processes that distinguish vascular trees more conspicuously from background. (2) The responses of the elongated representation encode the intrinsic profile of the vessels and capillaries, elongated-like shape, which retains the subtle intensity changes of vessels and capillaries for solving the continuity issue. (3) The proposed method disambiguates the region surrounding vessel structures for addressing the disturbance of noise at vascular regions. (4) Our method achieves the best quantitative and qualitative results over the state-of-the-art vessel delineation benchmarks.

back to top

A Modulatory Elongated Model for Delineating Retinal Microvasculature in OCTA Images