Pith. sign in

REVIEW 23 references

Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality

T0 review · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Siamese networks on an OCT foundation model detect nAMD change between visits, and a Wasserstein-distance loss that respects class ordering forecasts change three months out — both rank in the top four of the MICCAI MARIO challenge.

desk verdict Competent challenge report where the T1 Siamese work is solid and the T2 prediction claims overstep what 0.485 balanced accuracy supports. read the letter →

arxiv 2501.14323 v1 pith:ZTMVJ2IE submitted 2025-01-24 eess.IV cs.CVcs.LG

Taha Emre , Teresa Araújo , Marzieh Oghbaie , Dmitrii Lachinov , Guilherme Aresta , Hrvoje Bogunović This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords neovascularage-relatedmaculardegenerationopticalcoherencetomographySiamesenetworkordinalclassificationEarthMover'sdistanceWassersteinchangedetectiondiseaseprogressionprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper sets out to show that automated deep learning can cover both decision points in nAMD treatment monitoring from retinal OCT: telling whether disease activity changed between two visits, and forecasting whether it will change three months after a single visit. For the first, SiamRETFound, a Siamese network with shared weights on top of the RETFound OCT foundation model, compares two B-scans and reaches 0.713 balanced accuracy on the MARIO challenge's task-1 development set, ranking 4th. For the second, WARIO couples masked-autoencoder pretraining, focal loss, and the first Earth Mover's (Wasserstein) distance loss applied to forecasting disease-severity evolution, which treats Reduced, Stable, and Worsened as an ordered scale, ranking 2nd on task 2 with balanced accuracy 0.485 and rank correlation 0.206. A sympathetic reader would care because detecting change and predicting future change is exactly the information clinicians need to decide whether and when to inject anti-VEGF, and the paper's methods do this from standard B-scans without manual feature engineering.

What carries the argument

The load-bearing pieces are two training-time mechanisms. SiamRETFound is a Siamese network with shared weights: two copies of the RETFound encoder, a 24-block Vision Transformer with embedding size 1024 pretrained on retinal OCT by masked autoencoding, each embed one B-scan; the two embeddings are concatenated and a two-layer head outputs the change class, so the network learns which feature differences between visits are clinically meaningful. WARIO's load-bearing piece is an ordinal loss built from focal loss plus a discrete Earth Mover's (Wasserstein) distance term, $\ell_{\mathrm{EMD}} = \left( \frac{1}{C} \sum_{i=0}^{C-1} |\mathrm{CDF}_y(i) - \mathrm{CDF}_{\hat{p}}(i)|^2 \right)^{1/2}$, which is the $\ell^2$ distance between the cumulative distribution functions of the target and predicted class probabilities and therefore encodes the ordering Reduced < Stable < Worsened directly into the training signal. Around that loss sits masked-autoencoder pretraining at 75% masking, class-balanced finetuning, a 3-fold ensemble, and a volume-level postprocessing rule (Stable only if at least 80% of B-scans vote Stable, else majority vote) that the paper credits with the largest improvement in the ordinal Rk-correlation metric.

What would settle it

Retrain WARIO on the same MARIO task-2 data with the ordinal EMD term removed — cross-entropy and focal loss only, keeping the same MAE pretraining, ensembling, and postprocessing — and compare Rk-correlation on the official validation set; if the reported 0.206 does not fall, the ordinal loss is not what carries the forecast. A second check is to split the data patient-wise into disjoint train and validation folds instead of pooling B-scans as i.i.d. samples; if volume-level metrics collapse under that split, the patient-level correlation across B-scans was the source of the apparent skill.

Watch

Extended reading notes

Core claim

The paper's central claim is that both halves of nAMD monitoring can be automated from OCT B-scans, and that each requires the right inductive bias. For change detection, SiamRETFound compares two time-point B-scans through a shared-weight RETFound encoder, concatenates the embeddings, and classifies the pair as Reduced, Stable, Worsened, or Other; after pretraining on 300,000 surrogate change/no-change pairs from the public Kermany dataset and finetuning on MARIO task 1, it reaches 0.713 balanced accuracy as a single model and 0.801 average challenge score as a ten-model ensemble (5 SiamRETFound variants plus 5 Swin-transformer models), ranking 4th on the development leaderboard. For three-month forecasting, WARIO trains a Vision Transformer with masked-autoencoder pretraining and finetunes it with a loss that adds focal weighting to a squared Earth Mover's distance between the cumulative distributions of predictions and one-hot targets, so that mistaking Worsened for Reduced is penalized more than mistaking Worsened for Stable; with 3-fold ensembling and a volume-level postprocessing rule that labels a scan volume Stable only if at least 80% of its B-scans agree, and otherwise takes the majority vote, WARIO reaches 0.485 balanced accuracy and 0.206 Rk-correlation, ranking 2nd. The paper further reports that the focal-plus-EMD combination is strictly necessary, since without it the network collapses to predicting only the majority Stable class.

Load-bearing premise

The load-bearing premise is that 2D B-scans can be treated as independent training samples, even when many come from the same patient's OCT volume, and that those individual slices carry enough signal to predict the patient's volume-level activity change three months later, with the official labels trusted as ground truth.

Editorial extensions

If this is right

  • A Siamese change detector built on a retinal foundation model can spot subtle between-visit fluid changes that drive treatment decisions, not just gross deterioration.
  • Encoding the Reduced-Stable-Worsened ordering as a Wasserstein loss keeps the forecaster from collapsing onto the majority Stable class, the failure mode the authors observe without it.
  • Enforcing volume-level consistency in postprocessing improves the ordinal Rk-correlation more than any other ingredient, matching how clinicians actually score an eye.
  • Both tasks depend on strong pretraining, so foundation-model and masked-autoencoder representations are the transferable backbone for OCT-based nAMD monitoring.
  • Held-out leaderboard results (4th on task 1, 2nd on task 2) indicate the approach generalizes to the challenge's validation volumes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the focal-plus-EMD loss recipe is a general tool for ordinal medical classification — diabetic retinopathy grading, glaucoma staging, osteoarthritis severity — wherever classes are ordered and the majority class dominates; the paper demonstrates it on nAMD, but the mechanism is task-agnostic.
  • My inference: the 80%-Stable postprocessing threshold is a prior tuned to the challenge's class ratio; a clinic with different revisit schedules would need to recalibrate it, and the paper provides no procedure for doing so.
  • My inference: the change detector and the forecaster could be chained into a treat-or-watch aid — flag activity change at the current visit and forecast change at the next — which the paper gestures at as clinical workflow support but does not implement.
  • My inference: because training pools B-scans as i.i.d. while labels are volume-level, the reported metrics probably reflect a mix of genuine per-eye signal and within-volume label leakage; patient-stratified evaluation would separate the two.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the methods are empirical challenge submissions evaluated against an external benchmark, and the self-citations are contextual rather than load-bearing.

full rationale

This paper is a challenge submission whose claims are empirical and are assessed on the external MARIO leaderboard, not derived from a chain of equations that reduce to the input. The T1 Siamese model finetunes the external RETFound pretrained encoder with cross-entropy on the challenge training pairs; the T2 model combines a focal loss with a standard EMD/Wasserstein loss and post-processes volume-level predictions. No equation is defined in terms of the target outcome, and no fitted parameter is renamed as a prediction. The self-citations (TINC, Morph-SSL) are invoked only to support the general usefulness of pretraining in AMD progression tasks; the actual pretraining uses MAE on the challenge data, so the central result does not depend on those papers' unverified conclusions. The EMD loss is taken from external prior work (Hou et al., Talebi and Milanfar), and calling it a first use for this specific task is a novelty claim rather than a circular derivation. A data-hygiene ambiguity exists in Section 2.2, where MAE pretraining is said to use 'a combined dataset of T1 and T2' immediately after the T2 validation split is described, without explicitly stating that validation volumes were excluded; this is a leakage risk, not a circularity, and it does not make the prediction equivalent to the input by construction. Overall, no circular step is identifiable in the paper's method or evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on several hand-chosen hyperparameters (focal loss alpha/gamma, equal loss weighting, the 80% Stable threshold) and domain assumptions about label quality and B-scan independence. No new physical or clinical entities are introduced. The EMD loss is a standard tool applied to a new task, not a fitted constant.

free parameters (3)
  • Focal loss hyperparameters alpha and gamma = not reported
    The focal loss in Eq. (1) depends on alpha and gamma, which are not given numerical values in the paper; they are implicitly tuned to address class imbalance and affect all T2 results.
  • Loss weighting between focal and EMD terms = equal (1:1)
    The final T2 loss is described as the combination of the two terms with equal weighting, a manual choice not justified by ablation.
  • Volume-level Stable threshold = 80%
    In postprocessing, a volume is labeled Stable only if at least 80% of B-scan predictions are Stable; this threshold is chosen based on the class ratio and materially changes the final predictions.
assumptions (4)
  • domain assumption B-scan-level i.i.d. assumption during T2 training
    Section 2.2 states each B-scan is treated as i.i.d. even from the same patient, ignoring patient-level and volume-level dependence while labels are volume-level.
  • domain assumption Other class corresponds to uninterpretable scans
    Section 1.4 infers the Other label is associated with noise, obscured regions, flipping, or poor alignment from qualitative assessment, with no objective criteria and some nuanced cases.
  • domain assumption Surrogate pretraining labels from dataset disease classes approximate clinical change
    Section 2.1 creates change/no-change labels for Kermany pairs by comparing disease classes, assuming same-class pairs are no-change and different-class pairs are change; this may not reflect true nAMD activity change.
  • standard math EMD loss definition is valid and appropriate for ordinal classification
    The discrete EMD loss in Eq. (2) is taken from Hou et al. and Talebi and Milanfar; the paper does not re-derive it, and the ordinal relation Reduced-Stable-Worsened is assumed from label definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality." pith.science (2026). https://pith.science/paper/ZTMVJ2IE

@misc{pith2026250114323,
  author       = {Pith},
  title        = {Pith review of: Automatic detection and prediction of nAMD activity change in retinal OCT using Siamese networks and Wasserstein Distance for ordinality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZTMVJ2IE}},
  note         = {Machine review of arXiv:2501.14323}
}
read the original abstract

Neovascular age-related macular degeneration (nAMD) is a leading cause of vision loss among older adults, where disease activity detection and progression prediction are critical for nAMD management in terms of timely drug administration and improving patient outcomes. Recent advancements in deep learning offer a promising solution for predicting changes in AMD from optical coherence tomography (OCT) retinal volumes. In this work, we proposed deep learning models for the two tasks of the public MARIO Challenge at MICCAI 2024, designed to detect and forecast changes in nAMD severity with longitudinal retinal OCT. For the first task, we employ a Vision Transformer (ViT) based Siamese Network to detect changes in AMD severity by comparing scan embeddings of a patient from different time points. To train a model to forecast the change after 3 months, we exploit, for the first time, an Earth Mover (Wasserstein) Distance-based loss to harness the ordinal relation within the severity change classes. Both models ranked high on the preliminary leaderboard, demonstrating that their predictive capabilities could facilitate nAMD treatment management.

Figures

Figures reproduced from arXiv: 2501.14323 by the authors.

Figure 1
Figure 1. SiamRETFound approach for longitudinal change detection in retinal OCT. The full siamese network is first pretrained on a public OCT dataset using a simulated binary change detection task: change/no change. The network is then finetuned on the MARIO training data for classification of the target classes. Datasets and Evaluation For training and evaluating our model we used the MARIO challenge task 1 development data… view at source ↗
Figure 2
Figure 2. Examples of SiamRETFound predictions (GT: ground truth, P: prediction). this consistency in the postprocessing step combined with an ensemble of 3-fold data split. Even though F1-score dropped slightly, WARIO improved along other metrics. It is important to highlight that postprocessing improved Rk-correlation the most which is generally used in imbalanced classification problems. At the end, WARIO ranked 2nd in the… view at source ↗
Figure 3
Figure 3. Confusion matrices for Longitudinal change detection (MARIO Task 1). ] Reduced Stable Worsened Predicted label Reduced Stable Worsened True label 11 340 0 50 2736 35 12 632 6 (a) Single model Reduced Stable Worsened Predicted label Reduced Stable Worsened True label 174 177 0 205 2464 152 61 532 57 (b) Ensemble [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Confusion matrices for AMD evolution prediction (MARIO Task 2) [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Occlusion map sensitivity. Map values are from 0 (blue) to 1 (red). Lower values indicate the occluded region impacted more the final prediction [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: MAE pretraining with masking and reconstruction [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 12 canonical work pages

  1. [1]

    MARIO challenge (8 2024),https://www.codabench.org/competitions/2852

    work page 2024

  2. [2]

    IEEE Transactions on Medical Imaging (2024)

    Chakravarty, A., Emre, T., Leingang, O., Riedl, S., Mai, J., Scholl, H.P., Sivaprasad, S., Rueckert, D., Lotery, A., Schmidt-Erfurth, U., et al.: Morph-ssl: Self-supervision with longitudinal morphing for forecasting amd progression from oct volumes. IEEE Transactions on Medical Imaging (2024)

    work page 2024

  3. [3]

    In: International Conference on Learning Representations (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)

    2021

  4. [4]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Emre, T., Chakravarty, A., Rivail, A., Riedl, S., Schmidt-Erfurth, U., Bogunović, H.: Tinc: temporally informed non-contrastive learning for disease progression modeling in retinal oct volumes. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 625–634. Springer (2022)

    work page 2022

  5. [5]

    Computers & Electrical Engineering81, 106532 (2020)

    Gholami, P., Roy, P., Parthasarathy, M.K., Lakshminarayanan, V.: Octid: Optical coherence tomography image database. Computers & Electrical Engineering81, 106532 (2020)

    2020

  6. [6]

    The Lancet 401(10386), 1459–1472 (2023)

    Guymer, R.H., Campbell, T.G.: Age-related macular degeneration. The Lancet 401(10386), 1459–1472 (2023)

    work page 2023

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000–16009 (2022)

    2022

  8. [8]

    arXiv preprint arXiv:1611.05916 (2016)

    Hou, L., Yu, C.P., Samaras, D.: Squared earth mover’s distance-based loss for training deep neural networks. arXiv preprint arXiv:1611.05916 (2016)

    arXiv 2016

Show all 23 references
  1. [9]

    Karwande, G., Mbakwe, A.B., Wu, J.T., Celi, L.A., Moradi, M., Lourentzou, I.: CheXRelNet: An Anatomy-Aware Model for Tracking Longitudinal Relation- ships Between Chest X-Rays. Lecture Notes in Computer Science (including sub- series Lecture Notes in Artificial Intelligence an...

    2022 doi

  2. [10]

    Cell172(5), 1122–1131 (2018)

    Kermany, D.S., Goldbaum, M., Cai, W., et al.: Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning. Cell172(5), 1122–1131 (2018). https://doi.org/10.1016/j.cell.2018.02.010 10 Emre et al

    2018 doi

  3. [11]

    Scientific Data 11(1), 365 (2024)

    Kulyabin, M., Zhdanov, A., Nikiforova, A., Stepichev, A., Kuznetsova, A., Ronkin, M., Borisov, V., Bogachev, A., Korotkich, S., Constable, P.A., et al.: Octdl: Optical coherence tomography dataset for image-based deep learning methods. Scientific Data 11(1), 365 (2024)

    2024

  4. [12]

    Radiology: Artifi- cial Intelligence2(4), 1–39 (2020).https://doi.org/10.1148/ryai.2020200079

    Li, M.D., Arun, N.T., Gidwani, M., Chang, K., Deng, F., Little, B.P., Mendoza, D.P., Lang, M., Lee, S.I., O’Shea, A., Parakh, A., Singh, P., Kalpathy-Cramer, J.: Automated assessment and tracking of COVID-19 pulmonary disease severity on chest radiographs using convolutional s...

    2020 doi

  5. [13]

    https://doi.org/10.1038/s41746-020-0255-1

    Li, M.D., Chang, K., Bearce, B., Chang, C.Y., Huang, A.J., Campbell, J.P., Brown, J.M., Singh, P., Hoebel, K.V., Erdoğmuş, D., Ioannidis, S., Palmer, W.E., Chiang, M.F.,Kalpathy-Cramer,J.:Siameseneuralnetworksforcontinuousdiseaseseverity evaluationandchangedetectioninmedicalim...

    2020 doi

  6. [14]

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection.In:ProceedingsoftheIEEEinternationalconferenceoncomputervision. pp. 2980–2988 (2017)

    2017

  7. [15]

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

    2021

  8. [16]

    Advances in Neural Information Processing Systems35, 9201–9216 (2022)

    Prabhushankar, M., Kokilepersaud, K., Logan, Y.y., Trejo Corona, S., AlRegib, G., Wykoff,C.:Olivesdataset:Ophthalmiclabelsforinvestigatingvisualeyesemantics. Advances in Neural Information Processing Systems35, 9201–9216 (2022)

    2022

  9. [17]

    In: Predictive Intelligence in Medicine: Second International Workshop, PRIME 2019, Held in Conjunction with MICCAI 2019, Shenzhen, China, October 13, 2019, Proceedings 2

    Rivail, A., Schmidt-Erfurth, U., Vogl, W.D., Waldstein, S.M., Riedl, S., Grechenig, C., Wu, Z., Bogunovic, H.: Modeling disease progression in retinal octs with lon- gitudinal self-supervised learning. In: Predictive Intelligence in Medicine: Second International Workshop, PRI...

    2019

  10. [18]

    International Journal of Computer Assisted Radiology and Surgery (2023)

    Rochman, S., Szeskin, A., Lederman, R., Sosna, J., Joskowicz, L.: Graph-based automatic detection and classification of lesion changes in pairs of CT studies for oncology follow-up. International Journal of Computer Assisted Radiology and Surgery (2023). https://doi.org/10.100...

    2023 doi

  11. [19]

    Progress in Retinal and eye Re- search 50, 1–24 (2016)

    Schmidt-Erfurth, U., Waldstein, S.M.: A paradigm shift in imaging biomarkers in neovascular age-related macular degeneration. Progress in Retinal and eye Re- search 50, 1–24 (2016)

    2016

  12. [20]

    IEEE transactions on image processing27(8), 3998–4011 (2018)

    Talebi, H., Milanfar, P.: Nima: Neural image assessment. IEEE transactions on image processing27(8), 3998–4011 (2018)

    2018

  13. [21]

    Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)12907 LNCS, 670–680 (2021)

    To, M.S., Sarno, I.G., Chong, C., Jenkinson, M., Carneiro, G.: Self-Supervised Lesion Change Detection and Localisation in Longitudinal Multiple Sclerosis Brain Imaging. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture...

    2021 doi

  14. [22]

    Nature Medicine26(6), 892–899 (2020)

    Yim, J., Chopra, R., Spitz, T., Winkens, J., Obika, A., Kelly, C., Askham, H., Lukic, M., Huemer, J., Fasler, K., et al.: Predicting conversion to wet age-related macular degeneration using deep learning. Nature Medicine26(6), 892–899 (2020)

    2020

  15. [23]

    Zhou, Y., Chia, M.A., Wagner, S.K., et al.: A foundation model for generalizable disease detection from retinal images. Nature622(7981), 156–163 (2023).https: //doi.org/10.1038/s41586-023-06555-x Title Suppressed Due to Excessive Length 11 A Appendix ] Reduced Stable Worsened ...

    2023 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.