Retinal vessel segmentation, the pixel-by-pixel outlining of arteries and veins in an image of the eye’s fundus, is a foundational step in computer-aided diagnosis of a range of eye diseases. When a segmentation algorithm performs well, downstream software can quantify vessel tortuosity, detect narrowing, spot avascular zones and flag early signs of disease for ophthalmologists. When it performs poorly, the consequences ripple outward: thin capillaries vanish from the output map, vessel edges blur, and clinically meaningful detail is lost precisely where disease tends to appear first. The Chinese team argues that this is exactly the blind spot of existing methods. Fundus images pose a hostile environment for segmentation networks because vessels vary enormously in caliber, from thick arcades near the optic disc to capillaries barely a pixel wide; they branch into intricate, tree-like structures; they sit against backgrounds with low contrast; and they are frequently obscured by lesions or other pathological interference.

The dominant architectural template in medical image segmentation for years has been the encoder-decoder design, descended from the fully convolutional network and its famous cousin U-Net. In such designs, the encoder progressively downsamples the image, trading spatial resolution for increasingly abstract feature representations, and the decoder then upsamples those features back to full resolution to produce a segmentation mask. The trouble, as the authors explain, is that downsampling is ruthless with high-frequency information. Fine vessels, delicate branch points and sharp edges are encoded in the high-frequency components of an image, and repeated pooling operations can discard them irretrievably. What survives in the deepest layers tends to describe large structures well but thin structures badly, and decoder upsampling then compounds the problem by smearing whatever detail remains.

WRSG-Net attacks this problem from three directions, each embodied in a dedicated module. The first is a Parametric Wavelet Refined Encoder, or PWREC. Rather than relying solely on conventional downsampling, this encoder employs learnable wavelet decomposition, a mathematical technique adapted from signal processing in which an image is separated into components at different frequency bands and spatial locations. The decisive innovation is that the wavelet transform is parametric, meaning the network learns the decomposition itself during training rather than applying a fixed transform. This allows the encoder to adaptively capture high-frequency vascular detail, the fine texture and sharp transitions that characterize vessel boundaries, and to preserve it as the image passes through successive stages. The extracted features are then refined through a dual-attention mechanism, which learns to weight the most informative channels and spatial locations, mitigating the loss of fine vessel detail during downsampling.

The second component, the Inter-Scale Information Propagation Module, or ISIPM, addresses a different failure mode. In a typical encoder-decoder network, features at different scales are produced at different layers but often remain loosely coupled, joined only at skip connections that transfer them to the decoder. The ISIPM establishes an explicit layer-wise information propagation flow, allowing hierarchical features, from coarse global context down to fine local detail, to be integrated continuously across the network’s depth. The result the authors report is enhanced consistency and continuity in the segmented vessels: fewer breaks in the vascular tree, fewer disconnected fragments, and a more faithful reconstruction of the branching topology that clinicians rely on for assessment.

The third component is the Semantic Guided Decoder, or SGDC. Decoders commonly operate somewhat blindly, upsampling features with limited awareness of what the overall segmentation should look like. The SGDC instead performs context-aware feature aggregation under decoder semantic guidance, using high-level semantic information to direct how low-level features are assembled into the final pixel-level prediction. This design serves two purposes described by the authors: it suppresses noise originating in the encoder, preventing spurious activations from being painted into the output as phantom vessels, and it alleviates the blurring that typically accompanies decoder upsampling, sharpening vessel edges in the final mask.

Together, the three modules form a coherent design philosophy. Where prior approaches have often bolted attention blocks or multi-scale tricks onto otherwise conventional architectures, WRSG-Net reconsiders the information flow end to end, asking at each stage what detail is at risk of being lost and how it can be carried forward. The wavelet encoder protects high-frequency content at the point of greatest danger, downsampling; the inter-scale propagation module keeps hierarchical features talking to one another; and the semantic-guided decoder ensures that the final reconstruction is informed by high-level context rather than left to interpolate in the dark.

Benchmark performance is where such architectural claims must ultimately be judged, and the team evaluated WRSG-Net on the three most widely used public datasets for retinal vessel segmentation: DRIVE, STARE and CHASE_DB1. These datasets, each manually annotated by experts, serve as the de facto standard battlegrounds for segmentation algorithms, and their differing imaging conditions and patient populations make consistent performance across all three a meaningful achievement. On DRIVE, the Digital Retinal Images for Vessel Extraction dataset, WRSG-Net achieved an F1-score of 85.17 percent, accuracy of 97.84 percent, sensitivity of 84.31 percent and specificity of 97.88 percent. On STARE, the Structured Analysis of the Retina dataset, it reached an F1-score of 90.11 percent, accuracy of 98.04 percent, sensitivity of 85.06 percent and specificity of 98.71 percent. On CHASE_DB1, drawn from the Child Heart and Health Study in England, it recorded an F1-score of 84.11 percent, accuracy of 97.12 percent, sensitivity of 83.14 percent and specificity of 99.27 percent.

These numbers reward close reading. Accuracy and specificity are comparatively easy to inflate in vessel segmentation because the vast majority of pixels in a fundus image belong to the background; a network that simply predicted “no vessel” everywhere would score well above 90 percent accuracy. The metrics that separate strong systems from weak ones are sensitivity, the proportion of true vessel pixels correctly identified, and the F1-score, the harmonic mean of precision and recall that penalizes both missed vessels and false detections. WRSG-Net’s figures on all three benchmarks are competitive with, and on several measures superior to, most state-of-the-art approaches, according to the authors’ comparisons. More striking than the aggregate numbers, however, is where the gains concentrate. The team reports that the segmentation results show obvious advantages in the identification of tiny blood vessels and the repair of blurred edge details, which is precisely the regime where clinical value and technical difficulty both peak.

The clinical motivation behind the work is substantial. Diabetic retinopathy remains a leading cause of preventable blindness in working-age adults worldwide, and its earliest vascular signatures, microaneurysms, capillary dropout, venous beading, are subtle and require expert interpretation. Hypertensive retinopathy similarly encodes the history of blood pressure damage in the caliber and appearance of retinal vessels. Screening programs increasingly depend on automated image analysis to triage the enormous volumes of fundus photographs generated in population-scale programs, and the reliability of those programs is bounded by the fidelity of the underlying segmentation. A network that preserves capillary-level detail and clean vessel edges offers a stronger foundation for quantifying vascular change over time, and the authors position WRSG-Net as providing a new feasible technical path for retinal vessel segmentation with application potential in auxiliary screening and diagnosis of fundus vascular diseases.

The study was conducted entirely on publicly released, fully anonymized image datasets, involving no healthcare intervention on human participants, and the authors acknowledge the research teams behind the DRIVE, STARE and CHASE_DB1 benchmarks for their role in building the open resources on which the field depends. The research was supported by the Central Guidance on Local Science and Technology Development Fund Project (24ZYCXCP0020). The corresponding author is Guangle Yao of Chengdu University of Technology. The paper was received on 23 April 2026, accepted on 28 August 2026, and published on 05 September 2026 under open access terms that permit non-commercial sharing and reuse with attribution.

The broader significance of the work may lie in its demonstration that ideas from classical signal processing still have transformative power when fused with modern deep learning. Wavelets, which dominated image compression and denoising research for decades, fell out of fashion as convolutional networks learned their own feature hierarchies. WRSG-Net’s results suggest that the old and the new are complementary rather than competing: the wavelet transform offers an analytically sound way to represent multi-scale frequency content, while learned parameters and attention let the network decide which of those components matter for the task at hand. For the millions of patients whose eyes are photographed each year in screening programs, the difference between a segmentation that loses capillaries and one that keeps them may ultimately translate into earlier detection, better triage and preserved sight.

The retina, the thin layer of light-sensitive tissue lining the back of the eye, is the only place in the human body where blood vessels can be observed directly and non-invasively. This unique window makes retinal vessel imaging enormously valuable clinically. Changes in the width, curvature, branching pattern, and density of these vessels are early fingerprints of systemic and ocular conditions, including diabetic retinopathy, hypertensive retinopathy, and glaucoma, all of which can progress silently for years before a patient notices any loss of vision. Because early detection is the single most effective defense against permanent visual impairment, ophthalmologists have long sought automated systems capable of measuring the vasculature accurately and consistently. The manual tracing of vessels, however, is tedious, time-consuming, and subject to significant variability between different human experts, a problem the research team set out to eliminate.

The new study, led by Guanghui Song of Ningbo Tech University together with Binhua He of Zhejiang Sci-Tech University and Yan Nie of Ningbo University, focuses on a particularly challenging type of retinal imaging: fluorescein angiography, or FA. In this procedure, a fluorescent dye is injected into the bloodstream, and as it circulates through the retina, a specialized camera captures sequences of images in which the vessels glow brightly against a dark background. FA offers a dynamic, high-contrast view of blood flow that standard color fundus photography cannot match, revealing leakage, non-perfusion areas, and vascular abnormalities that would otherwise remain hidden. Yet the very characteristics that make FA so informative also make it difficult for automated analysis. Images are often affected by uneven illumination, background fluorescence, noise, and wide variation in vessel calibers, from thick major arcades down to capillaries only a pixel or two across.

To cope with this complexity, the researchers turned to multiscale analysis, a strategy rooted in the mathematics of wavelets. Wavelet transforms decompose a signal or image into components at different scales and resolutions, much as a musical equalizer separates sound into bass, midrange, and treble. The stationary wavelet transform in particular is translation-invariant, meaning it does not shift its representation when the input shifts, a property that makes it well suited to detecting structures of different sizes without introducing artifacts. By applying this transform to the angiographic images, the team generated a set of feature maps that emphasized vessels at each characteristic width. Fine capillaries standing out at fine scales, while large vessels emerged at coarser ones. This preprocessing step effectively converts the raw image into a richer representation in which vessel-like structures are amplified and background clutter suppressed.

These multiscale features then feed into a fully complex multiscale neural network, an architecture designed so that its internal processing parallels the scale hierarchy established by the wavelet stage. The core design principle is adaptation to two sources of variability that plague retinal image analysis: the enormous range of vessel widths and the constantly changing orientation of vessels as they sweep across the curved surface of the retina. Because the network receives explicit scale-decomposed inputs, it can assign different weights to different scales depending on context, learning to interpret a faint thin structure as a capillary rather than noise, and to distinguish a thick vessel from an illumination artifact. The complex-valued formulation of the network adds another dimension of representational power, allowing phase information carried by the wavelet coefficients to be exploited rather than discarded, which helps the model respond to the directionality of vascular structures.

Training such a network on relatively limited sets of expert-annotated medical images is a notorious bottleneck in medical artificial intelligence. Overfitting, in which a model memorizes the quirks of its training data rather than learning generalizable patterns, is a constant danger. The team addressed this with a deliberate data augmentation strategy: rotation operations were applied at least once across the layers during the training phase. In effect, the network was repeatedly confronted with the same vascular anatomy presented at different orientations, forcing it to learn features that are genuinely rotation-tolerant rather than tied to the particular angle at which vessels happened to appear in the training images. This echoes a broader theme in modern deep learning research, where equivariance to geometric transformations is prized precisely because biological structures, like retinal vessels, can appear at arbitrary orientations in any given scan.

The performance of the resulting system was evaluated on three different datasets, and in each case the proposed method delivered better results than the current state-of-the-art techniques it was compared against. This cross-dataset consistency matters greatly, because many published segmentation algorithms perform impressively on the specific dataset they were tuned to but degrade sharply when confronted with images from a different camera, a different patient population, or a different imaging protocol. Robustness across datasets is therefore one of the most honest indicators of whether a method has genuine clinical potential or is merely exploiting statistical idiosyncrasies of one benchmark.

Equally significant is the stability of the method. The researchers report that the framework produces consistent results across different training datasets and across inter-rater variability, the well-documented phenomenon in which two human experts annotating the same image disagree on fine details, particularly around the thinnest vessels and at branch points. A diagnostic tool that fluctuates depending on which annotator labeled its training data is of limited clinical value. The ability of the new method to absorb and transcend this human disagreement suggests it has learned a representation of vascular anatomy that is closer to the underlying biological reality than to the idiosyncrasies of any single annotation. As the authors note, this means the method can be practically used anywhere, a claim of portability that, if validated in prospective clinical studies, would be a meaningful advance for screening programs in settings where expert graders are scarce.

The clinical implications extend well beyond the technical achievement of drawing cleaner vessel maps. In diseases such as diabetic retinopathy, which affects a substantial fraction of the world’s growing diabetic population, the earliest signs of pathology are vascular: microaneurysms, capillary dropout, and changes in vessel tortuosity. Automated and accurate vessel segmentation is the foundational step upon which all such quantitative measurements depend. Fluorescein angiography is particularly central to assessing retinal perfusion, and recent work from other groups has explored using deep learning even to synthesize FA-like information from color fundus photographs. The present study complements that line of research by tackling the segmentation problem directly on authentic angiographic images, where the fluorescent signal provides a rich but noisy substrate for machine analysis.

The methodology also illustrates a persuasive middle path in medical AI design. Rather than relying solely on a generic convolutional architecture and hoping that enough data will teach it everything, the team built explicit domain knowledge into the pipeline. The choice of the stationary wavelet transform encodes decades of signal-processing understanding about how structures of varying scale can be separated, and the multiscale network architecture mirrors the physical fact that retinal vessels span a wide range of calibers. This hybrid of classical mathematics and modern deep learning, sometimes described as physics-informed or knowledge-guided machine learning, is increasingly seen as the most promising route to systems that are accurate, data-efficient, and trustworthy, the three qualities that regulators and clinicians alike demand from medical technology.

The work, which was partially supported by the Ningbo “Science and Technology Innovation Yongjiang 2035” key technology breakthrough plan project, arrives at a moment when the burden of retinal disease is rising worldwide. Diabetic retinopathy alone remains a leading cause of preventable blindness among working-age adults, and hypertension-related vascular changes in the retina are increasingly recognized as markers of broader cardiovascular risk. Screening programs that could deploy a stable, dataset-agnostic segmentation algorithm on angiographic images would multiply the reach of limited ophthalmological expertise, flagging patients who need urgent attention while sparing healthy individuals unnecessary interventions.

The authors caution, as with any new method, that broader clinical validation will be needed before deployment in routine care, and the study’s three-dataset evaluation, while encouraging, represents a step on a longer road toward regulatory approval and integration into hospital workflows. Nevertheless, the combination of superior performance, robustness across datasets, and stability against human annotation variability marks this framework as a notable contribution to the rapidly evolving field of retinal image analysis. If the promise holds, the humble wavelet, a mathematical tool first formalized more than a century ago in the work on orthogonal function systems, may find itself at the heart of software that safeguards the eyesight of millions.

Retinal foundation models represent a new generation of AI tools that can serve a wide range of ophthalmic applications—from automated disease diagnosis to prognosis and treatment response prediction. These models are typically “pre-trained” on large-scale datasets to imbibe generalizable image representations before being fine-tuned for specific tasks. Yet, despite their growing prominence, little was previously understood about how the nature of pre-training data impacts the model’s learned features, robustness, and clinical utility. Zhou and colleagues systematically tackled this knowledge gap by analyzing diverse pre-training scenarios using fundus image data from two geographically and demographically distinct cohorts.

The first fundus cohort in this study consists of over 100,000 images collected from a large urban hospital system, reflecting a broad spectrum of retinal pathologies, image qualities, and patient ethnicities. The second cohort, equally expansive, includes nearly 90,000 images obtained from a rural healthcare network, representing different socioeconomic and clinical contexts. By juxtaposing these datasets, the researchers could probe how medical, demographic, and imaging heterogeneity affects model behaviors in meaningful ways that previous work had not explored at this scale.

Central to their methodology was the construction of multiple foundation models pre-trained with varying subsets of these datasets, ranging from exclusively urban data to fully mixed urban-rural compositions. The team employed state-of-the-art convolutional neural network architectures tailored for high-resolution fundus images, meticulously optimizing training protocols to isolate the effects of pre-training data diversity and distribution. Subsequent evaluation of these models on independent diagnostic tasks revealed pronounced differences in performance metrics, particularly in sensitivity and specificity for diabetic retinopathy and glaucoma detection.

One of the most striking findings was that models pre-trained on more heterogeneous datasets, encompassing variations in ethnicity, disease prevalence, and imaging device characteristics, demonstrated superior generalizability when deployed on external test sets. This directly challenges the prevailing practice in AI ophthalmology of relying heavily on narrowly sourced images for pre-training, highlighting a tangible risk of model bias and reduced applicability in underrepresented patient subgroups. Zhou and colleagues’ results suggest that embracing data diversity at the pre-training stage not only bolsters accuracy but may also enhance health equity by minimizing disparities in AI-driven diagnoses.

The study further delved into feature representation analysis using advanced explainability tools to decode what the models learned during pre-training. Models trained on more diverse cohorts exhibited richer and more nuanced feature extraction capabilities, capturing subtle retinal texture variations and vascular patterns linked to early disease stages. In contrast, less diverse pre-training datasets yielded models inclined to overfit superficial image traits, thereby limiting their adaptability and clinical relevance. This highlights the intricate interplay between data heterogeneity and the learned internal representations that underpin successful deep learning models in ophthalmology.

Beyond performance metrics, the research team addressed practical considerations surrounding computational efficiency and data access constraints, which commonly influence dataset selection in clinical AI projects. By systematically evaluating model training time and convergence behavior relative to dataset size and diversity, they provide actionable guidance for balancing resource demands with model robustness. Their work advocates for collaborative data sharing and pooling strategies, particularly across heterogeneous cohorts, to accelerate the development of more reliable retinal AI tools.

The implications of this research extend beyond retinal imaging into broader medical AI domains, where the principles of foundation model pre-training and the impact of data provenance remain under-examined. Zhou et al.’s pioneering approach exemplifies how leveraging large-scale heterogeneous medical datasets can uncover latent biases and drive development of AI models that are both powerful and equitable. Given the rapidly increasing adoption of AI in clinical workflows, these insights are poised to influence regulatory considerations and best practices for dataset curation and model validation.

Furthermore, their investigation into the transfer learning paradigms prevalent in retinal AI effectively bridges engineering and clinical perspectives by demonstrating how foundational data choices ripple through to downstream diagnostic outcomes. This translational relevance makes the study a critical reference point for clinicians, AI developers, and healthcare policymakers seeking to harness AI’s full potential for eye health worldwide.

The authors also acknowledge limitations inherent in their approach, including the need for even broader population-level data encompassing additional geographic regions, and prospective clinical validation to assess model performance in real-world screening and diagnosis scenarios. Nonetheless, the scale and rigor of their work set a new benchmark in the ophthalmic AI research landscape and catalyze future studies aimed at refining dataset strategies to optimize foundation models for diverse clinical environments.

In conclusion, this seminal study reshapes our understanding of the pivotal role played by pre-training data in shaping retinal foundation models. By harnessing two vast and distinct fundus image cohorts, Zhou and colleagues have illuminated how data heterogeneity underpins model robustness, fairness, and clinical utility in profound ways. The findings encourage the AI in ophthalmology community to rethink data collection paradigms, prioritize inclusivity in dataset compilation, and rigorously evaluate pre-training effects—a paradigm shift that holds promise for advancing precision eye care globally through intelligent, equitable AI.

As AI-driven retinal diagnostics continue their rapid ascent, the lessons distilled from this research echo across broader medical imaging fields striving toward truly generalizable and unbiased artificial intelligence systems. Zhou et al.’s work stands as a clarion call to embrace data diversity not as an afterthought but as a foundational design principle—one that ultimately empowers AI to better serve the millions affected by vision-threatening diseases worldwide.

Subject of Research: The impact of pre-training data composition on the performance and generalizability of retinal foundation models using large-scale fundus image cohorts.

Article Title: Understanding pre-training data effects in retinal foundation models using two large fundus cohorts.

Article References:
Zhou, Y., Wang, Z., Wu, Y. et al. Understanding pre-training data effects in retinal foundation models using two large fundus cohorts. Nat Commun (2026). https://doi.org/10.1038/s41467-026-70077-z

Image Credits: AI Generated