An Explainable Five-Stage Pipeline: Estimating Arteriovenous Ratio from Fundus Images for Retinopathy Screening

A five-stage explainable pipeline outputs intermediate results at each stage for traceable fundus AVR estimation and screening.
This article presents a five-stage explainable pipeline for estimating the arteriovenous diameter ratio (AVR) from fundus images, breaking the process into vessel segmentation and arteriovenous classification, optic disc localization, minimum-chord boundary walking for caliber estimation, AVR computation of dominant branches, and explicit threshold-based screening decisions. Each stage outputs inspectable intermediate results so screening errors can be traced to specific steps. On 30 HRF images it reports 96.7% accuracy, 0.968 F1, and 0.978 ROC-AUC, but the study acknowledges limitations from the small sample, proxy disease labels, and lack of independent reference caliber measurements, requiring validation on larger clinically graded cohorts. Its core value is a transparent, debuggable research framework rather than clinical-grade screening performance.
Fundus images carry a wealth of information about vascular health, and the arteriovenous diameter ratio (AVR) is clinically regarded as a hallmark indicator of retinal vascular status. A recent study proposed a five-stage explainable pipeline that not only estimates AVR from fundus images but also outputs intermediate results at each stage, making screening decisions traceable.
Why "Explainable" AVR Estimation Is Needed
AVR reflects structural vascular changes by comparing arterial and venous diameters. Traditionally this relies on manual measurement or black-box models—the former is inefficient, while the latter makes it hard to pinpoint the source of errors. The core idea of this pipeline is to decompose the entire process into five independently inspectable stages: vessel mask and arteriovenous classification, optic disc localization, vessel caliber estimation based on minimum-chord boundary walking, AVR computation of dominant arteriovenous branches, and screening decisions based on explicit threshold rules.
Each stage outputs inspectable intermediate results, which means that when the final screening conclusion deviates, one can trace back to exactly which step went wrong rather than facing an inexplicable probability value.

Arteriovenous diameter ratio (AVR) is a classic quantitative metric in fundus image analysis, typically defined as the ratio of the central retinal artery equivalent diameter to the central retinal vein equivalent diameter. Clinical studies (such as ARIC and the Blue Mountains Eye Study) have found that a reduced AVR is associated with systemic vascular diseases including hypertension, diabetic retinopathy, and stroke risk. Traditional measurement follows standardized protocols such as Knudsen or Hubbard, measuring the calibers of major arteriovenous branches at a certain distance beyond the optic disc margin and then aggregating them according to a specific formula. Manual measurement is time-consuming and subject to inter-observer variability, making automated estimation a research hotspot; however, the outputs of black-box models are difficult to align with established clinical pathophysiological understanding, which is precisely the motivation for an explainable pipeline.
How the Five Stages Work Together
Vessel Segmentation and Arteriovenous Separation
The pipeline first extracts a vessel mask from the fundus image and further distinguishes arteries from veins. The output quality of this stage directly determines the reliability of subsequent caliber measurements. Unlike end-to-end models, the intermediate masks here can be visually inspected, allowing researchers to intuitively judge whether the classification is reasonable.
Vessel segmentation and arteriovenous classification are usually two coupled but distinct tasks. Vessel segmentation commonly uses semantic segmentation networks such as U-Net to extract a binary vessel mask from fundus images; arteriovenous classification requires further distinguishing arteries from veins, with common methods including color-channel-based features (arteries are usually brighter and redder, veins darker and bluer), rules based on vessel topology and optic disc location, or training dedicated classification networks. Classification errors propagate directly to AVR computation, because incorrectly merging arterial and venous calibers systematically shifts the ratio. Therefore, visual inspection at this stage is particularly critical in an explainable pipeline.
Optic Disc Localization and Caliber Measurement
Optic disc localization provides a spatial reference for vessel caliber measurement. Caliber estimation employs a "minimum-chord boundary walking" method—finding the shortest chord length along the vessel boundary to infer the diameter. This method is relatively robust to boundary noise, and its computation is transparent, without relying on a regression head trained on large-scale annotated data.
AVR Computation and Threshold Screening
AVR is derived by aggregating the calibers of the dominant arterial branch and the dominant venous branch. The screening decision does not come from a learned classifier but from an explicit threshold rule. This design sacrifices some flexibility in exchange for complete transparency of decision logic: given an AVR value and a threshold, the screening result can be reproduced by anyone.
Explicit threshold rules mean that screening decisions do not rely on a learned classifier but instead use a preset AVR cutoff value (for example, below a certain threshold is judged abnormal). The advantage of this approach is that the decision boundary is fully transparent and reproducible, and easy to align with reference thresholds in clinical literature; the disadvantage is that thresholds often depend on population and equipment, and may require recalibration when transferred across datasets. In contrast, end-to-end models can fit more complex decision boundaries but make it difficult to explain why a particular case was judged high-risk. In exploratory research, explicit thresholds help separate two sources of error: "measurement error" and "decision rule."
What the Exploratory Experimental Data Show
On an exploratory cohort of 30 HRF images, the pipeline reported 96.7% accuracy, an F1 score of 0.968, and a ROC-AUC of 0.978. These numbers are superficially impressive but need to be understood in the context of the experimental scale.
The study also conducted ablation experiments examining the impact of vessel pairing methods, caliber measurement methods, and aggregation strategies on the results, and made an exploratory comparison with VascX. The value of ablation experiments lies in revealing each stage's contribution to final performance, which is precisely the methodological advantage of an explainable pipeline over end-to-end models.
Interpretation Boundaries of Performance Metrics
The 96.7% accuracy is based on only 30 images, and the disease labels are proxy labels rather than clinical gold-standard grading. The study itself explicitly notes that interpretation is limited by the small cohort, proxy labels, and the lack of independent reference caliber measurements. In other words, these metrics demonstrate that the pipeline is feasible in exploratory settings, not that it achieves clinically usable screening performance.
The HRF dataset is a commonly used public dataset in the retinal vessel segmentation field, containing high-resolution fundus images and expert-annotated vessel masks, with images covering categories such as healthy, diabetic retinopathy, and glaucoma. Its scale is typically a few dozen images, suitable for methodological validation rather than clinical performance evaluation. Proxy labels refer to disease labels used in research that are not directly produced by clinical gold-standard grading but come from existing dataset classifications or surrogate criteria; this introduces label noise that can cause metrics such as accuracy to be overestimated or underestimated. ROC-AUC measures a model's ability to distinguish positive from negative samples across different thresholds; with class imbalance or small samples the variance is large, and single-point estimates should be interpreted with caution.
The Value and Limitations of a Transparent Framework
The real contribution of this pipeline is not in topping some benchmark but in providing a transparent framework for studying retinal vascular biomarkers. When AVR estimation goes wrong, researchers can investigate stage by stage: Did vessel segmentation miss small branches? Is arteriovenous classification confused? Or is caliber measurement inaccurate at crossing points? This debuggability is particularly important for biomarker research.
The limitations are equally clear. A small-sample cohort means limited statistical confidence, proxy disease labels may introduce label noise, and the lack of independent reference caliber measurements makes it impossible to verify the absolute accuracy of caliber estimation. The study's conclusions also acknowledge the need for further validation on larger, clinically graded cohorts.
For researchers concerned with explainability in medical imaging, the lesson of this five-stage design is that decomposing the computation of complex biomarkers into inspectable steps may have more long-term value than pursuing the accuracy of a single black-box model.

One-line Summary
A five-stage explainable pipeline outputs intermediate results stage by stage, enabling traceable fundus AVR estimation and screening.
Paragraph Summary
This article introduces a five-stage explainable pipeline for estimating the arteriovenous diameter ratio (AVR) from fundus images, decomposing the process into vessel segmentation and arteriovenous classification, optic disc localization, caliber estimation via minimum-chord boundary walking, AVR computation of dominant branches, and explicit threshold-based screening decisions. Each stage outputs inspectable intermediate results, allowing screening errors to be traced back to specific processing steps. On an exploratory cohort of 30 HRF images, the pipeline reported 96.7% accuracy, 0.968 F1, and 0.978 ROC-AUC, but the study itself notes limitations from the small sample, proxy disease labels, and the lack of independent reference caliber measurements, requiring validation on larger clinically graded cohorts. Its core value lies in providing a transparent, debuggable research framework rather than clinical-grade screening performance.
Related articles

Anthropic's IPO Eve: Five Risks Behind the Trillion-Dollar Valuation and the Truth About Its Governance
Breaking down Anthropic's IPO prospectus: $184M daily revenue at a trillion-dollar valuation, the truth behind $104B in paper losses, a five-class share governance structure, unusual disclosures in 94 pages of risk factors, and how the lock-up schedule may affect retail investors.

FBI Warns: Hackers Weaponize AI Agent ARTEX to Attack 200+ Organizations
FBI warns that AI penetration framework ARTEX is being used to attack 200+ organizations, with CrowdStrike tracking the VaporPanda group. We break down ARTEX's agent orchestration architecture, the real attack chain, loss data, and the FBI's five defense recommendations.

FBI Warns: Free AI Hacking Tool Artex Spirals Out of Control, Breaches Cloud Systems in Three Hours
FBI warns that the free open-source AI penetration tool Artex is being abused by criminals and state-backed hackers to breach cloud systems in as little as three hours. This article analyzes real attack cases, overlooked side-door risks, and defense measures enterprises and individuals can take immediately.