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.

A new artificial-intelligence system could turn the branching blood vessels visible at the back of the eye into a more precise source of medical information. Called MATHFI, the system analyzes retinal photographs and separates blood vessels from surrounding tissue, including the finest branches that are often difficult for automated methods to detect. In a study published in the Journal of Medical and Biological Engineering, researchers report that the network preserved the structure and continuity of retinal vessels while reducing false signals caused by the optic disc, bright lesions, uneven illumination, and other anatomical features. The goal is not simply to create a cleaner image. By producing a more faithful map of the retinal circulation, MATHFI is designed to support quantitative measurements that may help characterize diabetic retinopathy and other disorders linked to vascular damage.

The retina offers a rare view of the body’s microcirculation without the need for surgery. Light entering the eye is focused onto a layer of neural tissue supplied by a dense network of arteries, veins, and capillaries. In a standard fundus photograph, these vessels appear as dark, branching structures against a brighter background. Their arrangement, width, curvature, density, and branching complexity can change with disease. Yet extracting those features computationally is challenging. Thin vessels may occupy only a few pixels, contrast can vary across an image, and lesions or anatomical boundaries may resemble vessel segments. A model that incorrectly fills in background regions or breaks a vessel into disconnected pieces can distort downstream measurements, potentially making a healthy vascular network appear abnormal or masking a real change.

MATHFI addresses the problem by combining several image-processing and deep-learning strategies in one pipeline. Its name refers to Multi-scale Adaptive Thresholding with Hierarchical Feature Integration. Adaptive thresholding estimates whether a pixel belongs to a vessel relative to its local surroundings rather than applying one fixed brightness cutoff to the entire photograph. This is important because illumination is rarely uniform across a fundus image. The network also examines information at multiple spatial scales, allowing it to recognize both broad vessels and narrow peripheral branches. In effect, the model must solve two related problems at once: identifying the local visual signature of a vessel and understanding whether that candidate segment fits into the larger vascular architecture.

The system’s neural architecture uses attention-guided feature modulation, hierarchical skip fusion, multi-scale subtraction, adaptive selective fusion, and a final refinement stage. These terms describe ways of preserving useful information as an image passes through the network. In an encoder-decoder model, the encoder progressively compresses an image to learn increasingly abstract features, while the decoder reconstructs a pixel-level segmentation map. Skip connections transfer high-resolution details from early layers to later reconstruction stages, helping recover vessel boundaries that might otherwise be lost. MATHFI’s hierarchical fusion combines features from different depths rather than treating them as interchangeable. Attention mechanisms further assign greater weight to image regions and features that appear relevant to vessels, while suppressing patterns more likely to represent background anatomy or imaging artifacts.

Topology is central to the researchers’ approach. A vessel map is not merely a collection of correctly labeled pixels; it is also a network of connected paths. If a narrow vessel is interrupted by even a small gap, measurements of branching, tortuosity, or network density can change substantially. To assess this structural property, the study used clDice, a topology-aware metric that evaluates the overlap between the predicted and reference vessel skeletons as well as their regions. Conventional measures such as sensitivity and specificity count correctly classified pixels, but they may not adequately penalize a model that produces a visually plausible yet disconnected network. By emphasizing centerline continuity, clDice is better suited to the biological geometry of vessels, where connectivity matters for both interpretation and quantitative analysis.

The researchers trained and evaluated MATHFI on three publicly available retinal image datasets: DRIVE, CHASE-DB1, and STARE. These collections differ in image characteristics and acquisition conditions, making them useful benchmarks for testing whether a model can cope with variation beyond a single source. Using Monte Carlo cross-validation, the team repeatedly divided the data for training and evaluation rather than relying on one fixed split. On DRIVE, MATHFI achieved a sensitivity of 83.06 ± 0.92, a specificity of 98.09 ± 0.13, and a clDice score of 81.81 ± 0.36. On CHASE-DB1, the reported values were 85.00 ± 1.89 for sensitivity, 98.59 ± 0.24 for specificity, and 84.60 ± 0.66 for clDice. On STARE, the system reached 80.68 ± 8.49 sensitivity, 98.32 ± 0.61 specificity, and 82.65 ± 3.69 clDice. The variation in the results is itself informative: performance was less consistent on STARE, suggesting that image diversity remains a meaningful challenge.

The study then used MATHFI-generated vessel masks for a separate analysis of diabetic retinopathy. The researchers examined images from the APTOS 2019 Blindness Detection dataset, which includes five severity groups: no diabetic retinopathy, mild, moderate, severe, and proliferative disease. Rather than limiting the analysis to whether a vessel was present, they calculated four vascular biomarkers: fractal dimension, vascular density, tortuosity, and caliber. Fractal dimension is a mathematical description of how completely a branching pattern occupies space; in retinal imaging, it can capture aspects of the complexity of the vascular tree. Vascular density estimates the proportion or amount of vessel structure in a defined region. Tortuosity measures how much vessels deviate from a straighter course, while caliber describes their apparent width. Each measure provides a different view of retinal architecture, and each can be evaluated globally, within rings around the optic disc, or across the superior, inferior, nasal, and temporal regions represented by ISNT quadrants.

Across the diabetic retinopathy groups, the researchers found significant differences in all four biomarkers using one-way analysis of variance, with p values below 0.001. The reported global pattern included declining fractal dimension and vascular density as disease severity increased. Such a trend is biologically plausible because progressive retinal damage can involve capillary loss, nonperfusion, and remodeling of the vascular network, although the exact appearance of disease may differ between patients and anatomical regions. Changes in tortuosity and caliber can reflect vessel stress, abnormal remodeling, or altered blood flow, but these measurements are sensitive to image quality and segmentation errors. The researchers therefore present the biomarker results as evidence that the segmentation framework can enable quantitative analysis, not as proof that MATHFI alone can diagnose or stage an individual patient.

That distinction is important as artificial intelligence moves from experimental image analysis toward clinical decision support. A high-performing segmentation model does not automatically become a validated screening tool. It must be tested on diverse populations, cameras, image resolutions, ethnic groups, disease profiles, and real-world clinical workflows. It also needs to be compared with expert graders and evaluated for failure modes, including poor focus, media opacity, unusual anatomy, hemorrhages, exudates, and images captured outside the training distribution. The MATHFI study used established public datasets and reported an ablation analysis to examine the contributions of its major components, but the abstract does not establish prospective clinical performance or demonstrate that the system improves patient outcomes. Its immediate contribution is more foundational: a method for generating vessel maps that retain fine structures and support reproducible measurements.

If the approach withstands broader validation, its potential reach could extend beyond diabetic retinopathy. Retinal vascular patterns have been investigated in connection with hypertension, glaucoma, choroidal disease, and systemic vascular conditions. A reliable segmentation layer could allow researchers to measure subtle changes over time, compare vascular regions within the same eye, and combine vessel-derived features with lesion detection or other clinical signals. It could also make automated analysis more useful in settings where specialist graders are scarce, provided that appropriate safeguards and human oversight are maintained. For now, MATHFI is best understood as an enabling technology rather than a replacement for ophthalmic expertise. By treating the retina’s vessels as a connected biological network instead of a set of isolated dark pixels, the system points toward a future in which a routine eye photograph yields a richer, more quantitative portrait of disease.