pith. sign in

arxiv: 2604.10843 · v1 · submitted 2026-04-12 · 💻 cs.CV · cs.AI· cs.LG· cs.NE

Retinal Cyst Detection from Optical Coherence Tomography Images

Abhishek Dharmaratnakar , Aadheeshwar Vijayakumar , Suchand Dayanand This is my paper

Pith reviewed 2026-05-10 14:59 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LGcs.NE
keywords retinal cyst segmentationoptical coherence tomographyResNet CNNintraretinal cystsdice coefficientcystoid macular edemaimage segmentationmacular edema
0
0 comments X

The pith

ResNet CNN segments retinal cysts from OCT images with more than 70% Dice coefficient across all vendors.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that a ResNet convolutional neural network using patchwise classification can segment intraretinal cysts in optical coherence tomography scans more accurately than prior techniques. Earlier work reached only 68% accuracy and dropped further on noisy images from certain machine vendors. Training on the public cyst segmentation challenge dataset and evaluating on test images from four vendors annotated by two graders produces Dice scores above 70% in every case. Accurate cyst quantification supports better assessment of fluid buildup in diseases such as diabetic macular edema and age-related macular degeneration, which in turn improves predictions of visual acuity and guides clinical decisions.

Core claim

The authors establish that their ResNet CNN model, trained via patchwise classification on the cyst segmentation challenge dataset, delivers Dice coefficients exceeding 70% when segmenting intraretinal cysts in test OCT images from all four vendors, outperforming previous state-of-the-art methods limited to 68% accuracy while remaining robust to differences in image quality.

What carries the argument

ResNet CNN performing patchwise classification on OCT image patches to identify and segment cystic regions.

If this is right

  • More precise measurement of cyst volumes supports monitoring of cystoid macular edema progression.
  • Improved reliability in predicting visual acuity outcomes for patients with retinal fluid accumulation.
  • Reduced dependence on OCT image quality allows consistent results across different machine vendors in clinical settings.
  • The public challenge dataset enables standardized benchmarking for future automatic segmentation methods.
  • Lower variability in cyst assessment compared with differences between human graders.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The segmentation output could feed into automated systems that track changes in retinal fluid over multiple patient visits.
  • Extending the same patchwise approach to other OCT-detectable retinal conditions might expand its utility in comprehensive eye exams.
  • Larger-scale validation across additional clinical sites could confirm whether the reported performance holds for routine practice.
  • Volumetric analysis derived from the 2D segmentations may help quantify total fluid load more completely than area-based measures alone.

Load-bearing premise

The ResNet model trained on the challenge dataset will generalize to test images from different vendors and graders without being substantially affected by image quality variations or overfitting.

What would settle it

Dice coefficients falling below 70% when the trained model is applied to a new set of high-noise OCT images from an unseen vendor or with annotations from additional independent graders.

Figures

Figures reproduced from arXiv: 2604.10843 by Aadheeshwar Vijayakumar, Abhishek Dharmaratnakar, Suchand Dayanand.

Figure 1
Figure 1. Figure 1: Figure(above) represents 7 intra retinal layers with cyst in it. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Feedforward network 2) FeedBack ANN: Feedback loops are allowed. They are used in content addressable memories. It is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequenc… view at source ↗
Figure 3
Figure 3. Figure 3: Feedback network. generates a ”poor or undesired” output or an error, then the system alters the weights in order to improve subsequent results. 4) Back Propagation Algorithm: It is the training or learn￾ing algorithm. It learns by example. If you submit to the algorithm the example of what you want the network to do, it changes the network’s weights so that it can produce desired output for a particular i… view at source ↗
Figure 4
Figure 4. Figure 4: Residual Network information flow. The graphic above illustrates how the residual network achieves better information flow by passing more information through identity mapping to avoid going through the residual function. This ”network” acts just as a building block. In the ResNet architecture, the building block repeats many times to form a whole residual network. 2) ResNet General Architecture: Mathemati… view at source ↗
Figure 5
Figure 5. Figure 5: Feedforward Network vs Residual Network However, for the second network, there are two routes it can go, with one of them avoiding the weight matrix completely: ∂⃗y ∂h1 = ∂sigm ∂dot ∂dot ∂h1 + ∂sigm ∂h1 (6) To take the concept of information flow a step further, we can use a rectified linear unit instead of a sigmoid for our activation function [7]. If we use ReLUs and stack multiple layers together, any p… view at source ↗
Figure 6
Figure 6. Figure 6: Two actual classes or observed labels labels are used to compare with the predicted labels for performance evaluation after classification. In binary classification, a test dataset has two labels; positive and negative [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Predicted classes of a perfect classifier [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Predicted classes of a classifier The predicted labels of a classifier match with part of the observed labels. Confusion matrix from the four outcomes A confusion matrix is formed from the four outcomes produced as a result of binary classification. Four outcomes of classification A binary classifier predicts all data instances of a test dataset as either positive or negative [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 9
Figure 9. Figure 9: Four outcomes of a classifier [4]. This classification (or prediction) produces four outcomes true positive, true negative, false positive and false negative. • True positive (TP): correct positive prediction • False positive (FP): incorrect positive prediction • True negative (TN): correct negative prediction • False negative (FN): incorrect negative prediction Classification of a test dataset produces fo… view at source ↗
read the original abstract

Retinal Cysts are formed by leakage and accumulation of fluid in the retina due to the incompetence of retinal vasculature. These cystic spaces have significance in several ocular diseases such as age-related macular degeneration, diabetic macular edema, etc. Optical coherence tomography is one of the predominant diagnosing techniques for imaging retinal pathologies. Segmenting and quantification of intraretinal cysts plays the vital role in predicting visual acuity. In literature, several methods have been proposed for automatic segmentation of intraretinal cysts. As cystoid macular edema becomes a major problem to humankind, we need to quantify it accurately and operate it out, else it might cause many problems later on. Though research is being carried out in this area, not much of progress has been made and accuracy achieved so far is 68\% which is very less. Also, the methods depend on the quality of the image and give very low results for high noise images like topcon. This work uses ResNet CNN (Convolutional Neural Network) approach of segmentation by the way of patchwise classification for training on image set from cyst segmentation challenge dataset and testing on test data set given by 2 different graders for all 4 vendors in the challenge. It also compares these methods using first publicly available novel cyst segmentation challenge dataset. The methods were evaluated using quantitative measures to assess their robustness against the challenges of intraretinal cyst segmentation. The results are found to be better than the previous state of the art approaches giving more than 70\% dice coefficient on all vendors irrespective of their quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper proposes a ResNet CNN using patchwise classification to segment intraretinal cysts in OCT images. It trains on the cyst segmentation challenge dataset and evaluates on grader-annotated test data from four vendors, claiming Dice coefficients exceeding 70% across all vendors (including noisy Topcon scans) and outperforming prior state-of-the-art results of 68%.

Significance. If the performance claims are substantiated with proper methodology and reporting, the work would contribute to automated quantification of cysts, which is clinically relevant for monitoring diseases such as age-related macular degeneration and diabetic macular edema. The reliance on a public challenge dataset is a positive aspect for potential reproducibility.

major comments (3)
  1. Abstract: The headline claim of >70% Dice on all vendors irrespective of image quality lacks any supporting per-vendor breakdown, train/test split details (e.g., whether vendors are disjoint), or statistical significance tests. Without these, the modest improvement over 68% cannot be distinguished from dataset-specific effects or overfitting.
  2. Methods (implied in abstract): No description is provided of the patch extraction procedure, patch size, overlap, or how patchwise classifications are reconstructed into full-volume segmentations. This information is load-bearing for verifying the reported Dice scores and generalization.
  3. Abstract and evaluation: The text supplies no details on training hyperparameters, validation strategy, data augmentation, or handling of class imbalance, all of which are required to assess whether the ResNet generalizes to multi-vendor test data or simply fits the challenge training distribution.
minor comments (1)
  1. Abstract contains minor grammatical issues (e.g., 'giving more than 70% dice coefficient' and awkward phrasing around 'operate it out').

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important areas where the manuscript can be strengthened for clarity and reproducibility. We address each major comment below and will incorporate revisions to provide the requested details.

read point-by-point responses

  1. Referee: Abstract: The headline claim of >70% Dice on all vendors irrespective of image quality lacks any supporting per-vendor breakdown, train/test split details (e.g., whether vendors are disjoint), or statistical significance tests. Without these, the modest improvement over 68% cannot be distinguished from dataset-specific effects or overfitting.

    Authors: We agree that the abstract would be strengthened by including supporting details. In the revised manuscript, we will add a per-vendor breakdown of Dice coefficients (including for the noisy Topcon scans), explicitly describe the train/test splits from the cyst segmentation challenge dataset (with training on the public challenge data and testing on the multi-vendor grader-annotated sets), and report statistical significance tests or confidence intervals to substantiate the improvement over the prior 68% result. This will better demonstrate generalization across vendors. revision: yes

  2. Referee: Methods (implied in abstract): No description is provided of the patch extraction procedure, patch size, overlap, or how patchwise classifications are reconstructed into full-volume segmentations. This information is load-bearing for verifying the reported Dice scores and generalization.

    Authors: We acknowledge the need for these methodological details to ensure reproducibility. The revised Methods section will include a full description of the patch extraction procedure, the specific patch size employed, the overlap strategy between patches, and the reconstruction approach for combining patchwise classifications into full-volume segmentations (such as averaging or majority voting over overlaps). These additions will allow verification of the Dice scores and assessment of generalization. revision: yes

  3. Referee: Abstract and evaluation: The text supplies no details on training hyperparameters, validation strategy, data augmentation, or handling of class imbalance, all of which are required to assess whether the ResNet generalizes to multi-vendor test data or simply fits the challenge training distribution.

    Authors: We concur that these details are necessary to evaluate the approach. In the revision, we will expand the relevant sections to specify the training hyperparameters (e.g., optimizer, learning rate, epochs), validation strategy, data augmentation methods applied, and techniques for addressing class imbalance (e.g., weighted cross-entropy loss). This will clarify that the model generalizes to the multi-vendor test data rather than overfitting to the challenge training set. revision: yes

Circularity Check

0 steps flagged

Empirical ML evaluation with no derivation chain or self-referential steps

full rationale

The paper reports training a patchwise ResNet classifier on the cyst segmentation challenge dataset and measuring Dice coefficients (>70%) on grader-annotated test volumes from four vendors. No equations, first-principles derivations, fitted parameters renamed as predictions, or uniqueness theorems appear. The central claim is an empirical performance comparison against prior SOTA, not a logical reduction to its own inputs. No self-citations are load-bearing; the work is self-contained as a standard supervised segmentation experiment.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are introduced; the work relies on standard supervised learning with a pre-existing neural network architecture and public dataset.

pith-pipeline@v0.9.0 · 5592 in / 1202 out tokens · 64479 ms · 2026-05-10T14:59:11.219346+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 1 internal anchor

  1. [1]

    Houghton, and Amy L

    Wilkins, Gary R., Odette M. Houghton, and Amy L. Oldenburg. ”Auto- mated segmentation of intraretinal cystoid fluid in optical coherence to- mography. ”IEEE Transactions on Biomedical Engineering 59.4 (2012): 1109-1114

    work page 2012

  2. [3]

    Gonz ´alez, Ana, et al. ”Automatic cyst detection in OCT retinal images combining region flooding and texture analysis.” Proceedings of the 26th IEEE International Symposium on Computer-Based Medical Systems. IEEE, 2013

    work page 2013

  3. [4]

    ”Automatic cysts detection in optical coherence tomography images.” Mixed Design of Integrated Circuits & Systems (MIXDES), 2015 22nd International Conference

    Wieclawek, Wojciech. ”Automatic cysts detection in optical coherence tomography images.” Mixed Design of Integrated Circuits & Systems (MIXDES), 2015 22nd International Conference. IEEE, 2015

    work page 2015

  4. [5]

    ”Deep residual learning for image recognition.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    He, Kaiming, et al. ”Deep residual learning for image recognition.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016

    work page 2016

  5. [6]

    ”Segmentation of the optic disc in 3-D OCT scans of the optic nerve head.” IEEE transactions on medical imaging 29.1 (2010): 159-168

    Lee, Kyungmoo, et al. ”Segmentation of the optic disc in 3-D OCT scans of the optic nerve head.” IEEE transactions on medical imaging 29.1 (2010): 159-168

    work page 2010

  6. [7]

    ”Automatic segmentation of closed-contour features in ophthalmic images using graph theory and dynamic pro- gramming

    Chiu, Stephanie J., et al. ”Automatic segmentation of closed-contour features in ophthalmic images using graph theory and dynamic pro- gramming. ”Biomedical optics express 3.5 (2012): 1127-1140

    work page 2012

  7. [8]

    Roychowdhury, Sohini, et al. ”Automated localization of cysts in di- abetic macular edema using optical coherence tomography images.” 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2013

    work page 2013

  8. [9]

    ”Texture-based classification of focal liver lesions on MRI at 3.0 Tesla: A feasibility study in cysts and hemangiomas

    Mayerhoefer, Marius E., et al. ”Texture-based classification of focal liver lesions on MRI at 3.0 Tesla: A feasibility study in cysts and hemangiomas. ”Journal of Magnetic Resonance Imaging 32.2 (2010): 352-359

    work page 2010

  9. [10]

    ”Kernel regression based segmentation of optical coherence tomography images with diabetic macular edema.” Biomedical optics express 6.4 (2015): 1172-1194

    Chiu, Stephanie J., et al. ”Kernel regression based segmentation of optical coherence tomography images with diabetic macular edema.” Biomedical optics express 6.4 (2015): 1172-1194

    work page 2015

  10. [11]

    kellogg.umich.edu/patientcare/conditions/cystoid.macular.edema.html

    work page

  11. [12]

    eyewiki.aao.org/Cystoid Macular Edema

    work page

  12. [13]

    Abhishek, and K

    Vijayakumar, Aadheeshwar, D. Abhishek, and K. Chandrasekaran. ”DSL approach for development of gaming applications.” Information Systems Design and Intelligent Applications: Proceedings of Third International Conference INDIA 2016, V olume 1. Springer India New Delhi, 2016

    work page 2016

  13. [14]

    Abhishek, Aadheeshwar Vijayakumar, Pravin Bhaskar Ramteke, and Shashidhar G

    Chaugule, Vikrant, D. Abhishek, Aadheeshwar Vijayakumar, Pravin Bhaskar Ramteke, and Shashidhar G. Koolagudi. ”Product review based on optimized facial expression detection.” 2016 Ninth International Conference on Contemporary Computing (IC3). IEEE, 2016

    work page 2016

  14. [15]

    ”A semantic approach to text steganography in sanskrit using numerical encoding.” Recent Findings in Intelligent Computing Techniques: Proceedings of the 5th ICACNI 2017, V olume

    Vaishakh, K., et al. ”A semantic approach to text steganography in sanskrit using numerical encoding.” Recent Findings in Intelligent Computing Techniques: Proceedings of the 5th ICACNI 2017, V olume

    work page 2017

  15. [16]

    Springer Singapore, 2018

    work page 2018

  16. [17]

    ”Multi-Agent Video Recommenders: Evo- lution, Patterns, and Open Challenges.” arXiv preprint arXiv:2604.02211 (2026)

    Ranganathan, Srivaths, et al. ”Multi-Agent Video Recommenders: Evo- lution, Patterns, and Open Challenges.” arXiv preprint arXiv:2604.02211 (2026)

    work page arXiv 2026

  17. [18]

    ”Generative AI for Video Trailer Synthesis: From Extractive Heuristics to Autoregressive Creativity.” Authorea Preprints (2026)

    Dharmaratnakar, Abhishek, et al. ”Generative AI for Video Trailer Synthesis: From Extractive Heuristics to Autoregressive Creativity.” Authorea Preprints (2026)

    work page 2026

  18. [19]

    ”Factuality and Hallucinations in Large Language Models: A Comprehensive Survey.” Authorea Preprints (2026)

    Ranganathan, Srivaths, et al. ”Factuality and Hallucinations in Large Language Models: A Comprehensive Survey.” Authorea Preprints (2026)

    work page 2026

  19. [20]

    Beyond Fluency: Toward Reliable Trajectories in Agentic IR

    Sinha, Anushree, et al. ”Beyond Fluency: Toward Reliable Trajectories in Agentic IR.” arXiv preprint arXiv:2604.04269 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026