acceptodds
Under review as a conference paper at ICLR 2027

A General Bézier Tree Encoding Counterfactual Framework for Retinal-Vessel-Mediated Disease Analysis

Abstract

The geometry of the retinal vessel is a key biomarker of vascular diseases, yet clinical evidence remains primarily observational. Existing generative counterfactuals intervene only at the image-level disease label, failing to isolate explicit anatomical structure. To address this limitation, we propose the Bezier Tree Encoding Counterfactual Framework (BTECF). By abstracting vascular networks into interconnected cubic-Bezier segments, BTECF establishes a disease-agnostic representation in which structural topology is explicitly preserved and atomically perturbable. Coupling this encoding with a diffusion-based generator enables parameter-level do-interventions on explicit geometric axes (e.g., tortuosity, caliber) while preserving background fundus textures. We validate BTECF on diabetic retinopathy, together with independent cohorts for ischemic stroke and Alzheimer's disease. Isolated counterfactual interventions produce dose-responsive shifts in classifier predictions; a matched pixel-drop control attenuates this response by an order of magnitude or more, ruling out out-of-distribution generation artifacts. By enforcing causal isolation between vessel topology and pixel-level confounders, BTECF provides a unified generative paradigm for hypothesis verification across systemic diseases. To support reproducibility, the code will be publicly released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Synthesis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WAyvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes the Bézier Tree Encoding Counterfactual Framework for analyzing how retinal-disease classifiers respond to explicit modifications of vascular geometry. The framework first converts segmented retinal vascular trees into interconnected cubic Bézier segments, from which both a 20-dimensional geometric representation and a three-channel structural conditioning signal are derived. The Bézier representation is then coupled with a LoRA-adapted Stable Diffusion model and VesselControlNet to generate fundus images under controlled modifications of vessel tortuosity, arc length/segment presence, and caliber. Disease-specific classifiers are subsequently used to quantify prediction changes between baseline and perturbed generations.

Critical strengths
  1. The paper attempts to provide an explicit and editable representation of retinal vascular geometry.
  2. The authors conduct several sensitivity analyses and are relatively explicit about important limitations.
Critical weaknesses
  1. The central counterfactual and causal interpretation is not supported by the current methodology or experiments. The manuscript repeatedly uses terms such as “counterfactual,” “do-intervention,” and “ATE,” but the proposed procedure is essentially: modify a Bézier representation, pass the edited structure through a diffusion-generation pipeline, and measure the change in a classifier output. There is no explicit structural causal model, no identifiability argument, and no demonstration that the manipulated vessel variable is changed independently of other disease-relevant image characteristics. The manuscript itself acknowledges that geometric and image attributes may covary downstream after an intervention.
  2. I have substantial concerns about the scientific reasoning and presentation of the work. The manuscript explicitly states that generative AI tools were used for writing, literature retrieval, manuscript critique. The use of AI tools itself is not the issue. My concern is that the manuscript contains many signs of an AI-assisted style of scientific construction: a large number of formal terms, mathematical expressions, named modules, causal terminology, and layered technical descriptions are introduced, but the underlying principles and scientific motivation are often not developed with corresponding depth.
  3. The paper does not quantitatively verify that the generated images actually realize the intended geometric manipulation while preserving unrelated vascular properties, lesions, background appearance, and other classifier-relevant cues. The authors explicitly list remeasurement of target/non-target geometry, deformation-matched controls, direct mask/skeleton baselines, and blinded expert evaluation as future work.
Questions for authors

I made a serious effort to understand the paper’s reasoning, but my conclusion is that, in several places, mathematical and technical formalism appears to be used to package an engineering pipeline rather than to derive the method from a sufficiently clear understanding of the biological mechanism, causal assumptions, or intervention semantics. For example, labeling a Bézier perturbation followed by diffusion synthesis as a “do-intervention” does not make it a causal intervention unless the authors establish what variables are being held fixed, what assumptions justify the intervention, and why downstream changes can be attributed specifically to the edited vascular factor. The manuscript currently does not provide that level of reasoning.

Reviewer_p135OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Bézier Tree Encoding Counterfactual Framework (BTECF), which represents retinal vessels using cubic Bézier curves, applies geometric edits, and generates corresponding fundus images through conditional diffusion. Classifier predictions are compared between baseline and edited generations. Experiments focus on diabetic retinopathy (DR), with extensions to ischemic stroke and Alzheimer’s disease.

Critical strengths
  1. Meaningful research question. Examining classifier responses to explicitly specified anatomical changes is relevant to understanding their dependence on vascular structure.
  2. Coherent framework with initial empirical support. A shared vessel representation connects feature extraction, geometric editing, and image generation. The paper reports primary DR results for 500 starting images, alongside analyses of prompt conditions and classifier backbones.
Critical weaknesses
  1. The added value of the Bézier-based interface is insufficiently demonstrated. The paper presents a unified editing and analysis interface, but the evaluation does not establish what the Bézier representation contributes beyond enabling the pipeline. Comparisons with direct mask- or skeleton-based editing are absent, and the reported classifier responses do not establish an advantage in controllability or analytical utility. This leaves the significance of the proposed integration insufficiently supported.
  2. Intervention fidelity and response attribution lack essential validation. The evaluation reports classifier probability changes without quantifying the achieved geometric edits or accompanying changes in other vessel properties, background appearance, and lesions. The pixel-deletion comparison does not isolate these factors because it is not matched to geometric deformation. Consequently, the results establish responses to the editing pipeline but do not adequately identify responses to the targeted vascular attributes.
Questions for authors
  1. What measurable advantage or additional analytical capability does the Bézier-based interface provide over simpler editing representations under comparable generation settings?
  2. How accurately are the requested edits realized in generated images, and to what extent do accompanying non-target changes account for the classifier responses?
Reviewer_WiaXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper provide a Bezier Tree Encoding Counterfactual Framework (BTECF) for the Retinal-Vessel-Mediated disease analysis. They build a tool to audit disease classifiers on retinal (fundus) images: does the classifier actually respond to vessel geometry — tortuosity, arc length, vessel radius — or to something else? In this framework, each image's vessel network is fitted as a tree of cubic-Bézier curves, giving a single structured encoding of the vessel geometry. That encoding serves three roles at once: it provides an explicit set of geometric descriptors of the vessels; it conditions a diffusion generator that renders fundus images from it; and it is the object that gets edited. To probe a classifier, they change one geometric property in the encoding — for example making the vessels more tortuous — generate a paired baseline and edited image from the same starting point, and measure how the classifier's disease prediction changes between the two.

Critical strengths

This topic is quite interesting to me, which is also quite usful for the diease understanding. And it also general and can be extend to other imaging-related disease.

Critical weaknesses
  1. The method seems quite rely on the exsiting generative model, sure this is a critical problem. I have a high concern about the model for the counterfactual generation, does this model provide good enough reconstruction? does this model can recover your original input if you do a edit and then reverse that operation?

  2. For Stroke and the Alzheimer's, seems the dataset is too small thus the results may have a good deal of variance.

Questions for authors
  1. Like I mentioned above, how relaible is the iamge generated from the editing? One mroe thing we need to do is: Takes the generated image, segments its vessels, fits the tree again, and asks: did the vessels in this image actually become more tortuous? Did the arc length and radius stay where they were? Did the background, colour, and lesions stay the same?

  2. So for counterfactual generation, we can aslo do several things to ensure the mdoel is reliable: a. Null intervention, when we give a 0 edit to the model, will we get a perfect recontruction? I highly doubt about this. 2. what if you edit an image and reserse that operation based on the first first output? Can you get a perfect original image? What if you do this ten times?

  3. The classifier's DR probability jumps (Δ = +0.432), can this caused by the changes of other factors, for example the SNR of the image, and the subtle/invisible changes (to human eye) , not have to be the disease pattern?

In general I like the way the authors doing the study and the method of this paper, I am looking forward to the response.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.