⚡ electrical engineering

SpecF2M: A Spectral-Aware Multi-task Network Estimating Axial Length and Refractive Error from Pediatric Fundus Photographs

The paper introduces SpecF2M, a spectral-aware multi-task network that leverages a hybrid spatial-spectral backbone and anatomy-guided enhancement to accurately estimate pediatric axial length and refractive error from fundus photographs, achieving superior performance over baseline models with mean absolute errors of 0.5347 mm and 0.7062 D respectively.

Original authors: Mengxian He, Xinyue Liu, Yunyun Sun, Wei Hao, Minqing Zhang, Lichun Wang, Shunyi Zhang, Wu Yuan

Published 2026-08-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mengxian He, Xinyue Liu, Yunyun Sun, Wei Hao, Minqing Zhang, Lichun Wang, Shunyi Zhang, Wu Yuan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your eyes are like high-end cameras. Just as a photographer needs to know exactly how long the camera body is and how strong the lens needs to be to get a sharp picture, doctors need to know two specific things about your eyes to check if you are nearsighted (myopic). First, they need to measure the Axial Length, which is simply how long your eyeball is from front to back. If it grows too long, like a balloon stretching out, light focuses in front of your retina instead of on it, making distant things blurry. Second, they need to check your Refractive Error, which is the "power" of your eye's lens. This is usually split into a "Sphere" (the main power, like the main focus knob) and a "Cylinder" (which fixes astigmatism, or if your eye is shaped more like a football than a basketball).

Traditionally, getting these numbers is a bit of a hassle. Measuring the length of the eye requires special, expensive machines, and checking the lens power often involves putting strong, uncomfortable drops in a child's eyes to temporarily paralyze the focusing muscle. This makes it hard to screen millions of kids quickly. However, there is a simpler picture doctors already take: a fundus photograph. This is a photo of the back of the eye, like taking a picture of the inside of a cave. When a child's eye stretches to become nearsighted, the "cave" walls (the retina and blood vessels) get stretched and distorted in very specific ways. The big question scientists are asking is: Can we look at these stretched-out photos and use a computer to guess the eye's length and lens power, skipping the scary drops and big machines?

Enter SpecF2M, a new computer program designed to solve this puzzle. Think of SpecF2M as a super-smart detective that doesn't just look at a fundus photo; it listens to the "music" hidden inside the image. The researchers realized that the changes caused by nearsightedness aren't just about shapes; they are also about patterns and frequencies, much like how a song has both a melody and a rhythm. To catch these clues, SpecF2M uses a three-step trick. First, it acts like a photo editor that brightens up the faint, blurry details of the eye's back wall, making the stretched blood vessels and the optic nerve stand out clearly. Second, it uses a special "spectral" brain that can see both the tiny textures and the big, sweeping curves of the eye all at once, rather than just looking at one thing at a time. Finally, it uses a "team of experts" approach. Instead of one brain trying to guess everything, the system splits the work: one expert focuses on guessing the eye's length, another on the main lens power, and a third on the astigmatism, all while sharing what they learn from each other.

When the team tested this system on a huge group of over 4,000 children and nearly 7,000 eye photos, the results were promising. The program managed to guess the eye's length with an average error of just 0.5347 mm and the main lens power (Sphere) with an error of 0.7062 D. These numbers are better than other standard computer models they compared it against. However, the study also found something interesting and a bit tricky: while the system was great at guessing the eye's length and main power, it was less sure about the astigmatism (Cylinder). The data suggests that the stretching of the eye is tightly linked to the main power and length, but the "football shape" of astigmatism doesn't always leave a clear, consistent fingerprint on the fundus photo.

The authors are careful to say that while this is a huge step forward for screening, it isn't a magic wand yet. The system needs to be tested on even more groups of people from different places before it can be used in real clinics to replace the traditional methods. But for now, SpecF2M shows that by listening to the spectral "music" of the eye's back wall, we might soon be able to spot nearsightedness in kids just by looking at a simple, painless photo.

Technical Summary: SpecF2M

Problem Statement
Pediatric myopia is a growing global public health concern, characterized primarily by axial elongation of the eye. Clinical assessment relies on two core indicators: Axial Length (AL) and Spherical Equivalent Refraction (SER), the latter derived from Sphere (SPH) and Cylinder (CYL) components. However, obtaining these measurements typically requires cycloplegic refraction (time-consuming and uncomfortable for children) and dedicated biometry devices (e.g., Lenstar), limiting the scalability of large-scale screening.

While 45° posterior-pole fundus photography is a widely available, low-cost, and objective alternative, estimating AL and refraction from these images is challenging. Myopia-related changes in fundus images—such as optic disc (OD) deformation, peripapillary atrophy (PPA), fundus tessellation (FT), and vessel straightening—are often low-contrast, spatially diffuse, and multi-scale. Furthermore, existing deep learning approaches often rely on coarse joint predictions or standard CNN features, failing to adequately address the specific anatomical coupling between AL, SPH, and CYL, or to leverage the spectral nature of these structural changes.

Methodology: SpecF2M
The authors propose SpecF2M, a spectral-aware multi-task network designed to estimate AL, SPH, and CYL simultaneously from pediatric fundus photographs. The architecture consists of three primary modules:

Anatomy-Guided Enhancement Strategy:
To improve the discriminability of low-contrast myopic structures, the authors introduce a deterministic, training-free preprocessing operator $A(\cdot)$ applied to the input image $I$ . This operator comprises four steps:
- Luminance Enhancement: Conversion to LAB space followed by Contrast Limited Adaptive Histogram Equalization (CLAHE) on the luminance channel.
- High-Frequency Emphasis: Adaptive high-pass sharpening to highlight micro-structures while reducing low-frequency redundancy.
- Vessel-Prior Soft Suppression: Derivation of a soft vessel mask from the green channel using multi-orientation black-hat filtering, which is used to suppress linear vascular textures that may obscure diffuse tessellation patterns.
- OD/PPA ROI Enhancement: Deterministic morphological fitting to identify the optic disc region, followed by localized contrast redistribution and soft fusion to avoid boundary artifacts.
Spectral-Aware Representation Learning (Backbone):
The network employs a shared backbone that interleaves MixCNN blocks with Hybrid Spectral Learning (HSL) blocks to capture both spatial and spectral features.
- MixCNN Blocks: These combine standard convolutions (for stable local textures) with deformable convolutions (to adapt to geometric deformations in irregular anatomical structures).
- HSL Blocks: These extend representation into the spectral domain by cascading two components:
  - Wavelet-Selective CNN (WSConv): Performs multi-resolution wavelet decomposition, applying subband-specific convolutions, and uses a gating mechanism to modulate low-frequency responses based on high-frequency (LH) cues.
  - Wave Propagation Operator (WPO): A Fourier-domain global propagation module that evolves features over an internal time step using underdamped wave dynamics, capturing global frequency interactions.
Expert-Routing Multi-Task Learning:
Instead of a single shared head, SpecF2M utilizes an expert-routing mechanism to handle the multi-task nature of the problem.
- Three independent expert MLPs ( $g_\phi, g_\psi, g_\gamma$ ) transform the shared backbone features into task-specific embeddings for AL, SPH, and CYL.
- Task-specific gates predict routing weights to create a convex combination of these expert embeddings.
- This allows the model to balance shared anatomical representations with task-specific disentangling, addressing the observation that AL and SPH share stronger correlates than CYL.

Key Results
The model was evaluated on a dataset of 6,966 fundus images from 4,359 pediatric visits (mean age 8.22 years). SpecF2M was compared against standard CNN (ResNet18, DenseNet121, VGG16) and Vision Transformer (ViT) baselines.

Performance: SpecF2M achieved the lowest Mean Absolute Error (MAE) and highest $R^2$ $R^{2}$ for both AL and SPH estimation.
- AL: MAE = 0.5347 mm, $R^2$ = 0.6038.
- SPH: MAE = 0.7062 D, $R^2$ = 0.7251.
- CYL: MAE = 0.5508 D, $R^2$ = 0.4632.
Ablation Studies: Removing the anatomy-guided enhancement, the HSL blocks, or the expert-routing mechanism consistently degraded performance, particularly for AL and SPH.
Task Coupling Analysis: The results revealed asymmetric task coupling. AL and SPH showed a strong negative correlation (corr = -0.521) and benefited significantly from multi-task learning. In contrast, CYL showed near-zero correlation with AL (corr = -0.068) and exhibited weaker association with fundus-derived patterns, resulting in less stable estimation compared to the other metrics.

Significance and Claims
The paper claims that SpecF2M is the first framework to jointly estimate AL, SPH, and CYL from 45° pediatric fundus photographs using a spectral-aware, anatomy-guided approach. The authors highlight the following contributions:

Screening Utility: The method offers a potential pathway for accessible, image-based decision support for pediatric myopia screening, bypassing the need for immediate biometry or cycloplegia in initial stages.
Spectral Insight: By integrating wavelet and Fourier-domain propagation, the model explicitly captures the multi-scale and spectrally structured nature of myopic fundus changes, outperforming standard spatial-only baselines.
Component-Level Understanding: The study provides empirical evidence of the asymmetric coupling between myopia indicators, demonstrating that while AL and SPH are tightly linked to posterior-pole structural changes, CYL estimation remains more stochastic and less dependent on these specific fundus patterns.

The authors conclude that while SpecF2M outperforms controlled baselines, external validation is necessary before clinical deployment.

Technical Summary: SpecF2M

More like this