💻 computer science

HAG-EUNet: Attention-Guided Deep Learning for Choroid Segmentation and Clinical Biomarker Extraction in Optical Coherence Tomography

The paper introduces HAG-EUNet, a hybrid deep learning model combining an EfficientNet-B3 encoder with a transformer-enhanced U-Net decoder to achieve automated, high-precision choroidal segmentation in OCT images and extract 32 clinically relevant biomarkers, significantly outperforming standard U-Net baselines in both accuracy and efficiency.

Original authors: Jom Binoy, Gayathri S

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jom Binoy, Gayathri S

Original paper licensed under CC BY 4.0 (https://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

In the quiet, dark world inside the human eye, a thin, vascular layer called the choroid acts as a vital lifeline. It sits just beneath the retina, the light-sensitive tissue at the back of the eye, and its primary job is to deliver oxygen and nutrients to the outer retina. When this layer changes in thickness or shape, it often signals that something is wrong. These changes are linked to a range of serious eye conditions, from age-related macular degeneration to diabetic retinopathy, and recent research even suggests they may reflect broader health issues like heart disease. To understand these changes, doctors use a scanning technology called optical coherence tomography, or OCT. This device creates incredibly detailed cross-sectional images of the eye, much like an ultrasound but using light instead of sound. However, reading these images is difficult. The choroid often looks faint and blurry in the scans, obscured by a grainy texture known as speckle noise and shadowed by the blood vessels that run through it. For a doctor to measure the choroid accurately, they must manually trace its boundaries on the screen, a tedious process that can take twenty to twenty-five minutes for a single scan. This bottleneck has long prevented the rapid, large-scale analysis needed to catch diseases early.

A team of researchers at the Vellore Institute of Technology in India has developed a new approach to solve this problem, creating an automated system that can map the choroid with remarkable speed and precision. They built a specialized computer program named HAG-EUNet, which functions as a highly trained digital assistant for ophthalmologists. Instead of relying on a human to draw lines on every image, this system uses a deep learning model to analyze the OCT scan and automatically identify the exact boundaries of the choroidal layer. The researchers designed this model to be a hybrid, combining different advanced techniques to handle the specific challenges of eye imaging. It starts by cleaning up the noisy images, removing the grainy interference while keeping the delicate anatomical structures intact. Then, it uses a sophisticated architecture that acts like a pair of eyes with multiple lenses, capable of seeing both the fine details of the tissue and the broader context of the entire eye structure simultaneously.

The performance of this new system was tested against a variety of existing methods, including standard models that have been used for years. In a trial involving eighty-three test images that the model had never seen before, HAG-EUNet achieved a success rate of nearly 96 percent in correctly identifying the choroid. This score is significantly higher than the next best method, which managed about 95 percent, and far superior to older, more basic models that hovered around 91 to 93 percent. More importantly, the new system was far more consistent, producing results with very little variation from one image to the next. The researchers found that the system was particularly good at handling the difficult parts of the image, such as the edges where the choroid meets the white of the eye, a boundary that often confuses other software. By refining these edges, the model ensures that the measurements it takes are not just guesses but reliable data points.

Beyond simply drawing a line around the choroid, the system goes a step further to extract thirty-two different clinical measurements, or biomarkers, from the scan. These include the average thickness of the layer, the area it covers, and even complex descriptions of its texture and the pattern of its blood vessels. To verify that these automated numbers were trustworthy, the researchers compared them against measurements taken by a human expert ophthalmologist. The results showed a strong agreement between the machine and the human, with the automated measurements matching the expert's findings almost perfectly. This level of accuracy suggests that the system can be trusted to provide the same diagnostic insights that a specialist would, but in a fraction of the time. The researchers also demonstrated that the system could be integrated into a web application, allowing for real-time analysis where a doctor could upload a scan and receive a full report instantly.

The study highlights that while the new model is a significant step forward, it is not without its limits. The system was trained and tested on images from a single public dataset, and the researchers acknowledge that it may need further validation across different types of scanners and patient populations to ensure it works everywhere. Additionally, the current version analyzes flat, two-dimensional slices of the eye rather than the full three-dimensional volume, which could offer even more detail in the future. Despite these caveats, the work represents a tangible shift in how eye health is monitored. By automating the tedious task of segmentation, the system removes the burden of manual tracing and opens the door for faster, more consistent diagnosis of eye diseases. It transforms a slow, labor-intensive process into a rapid, automated workflow, potentially allowing doctors to detect subtle changes in the choroid before they lead to vision loss, and perhaps even to spot early signs of systemic health issues hidden within the eye.

Technical Summary: HAG-EUNet for Choroid Segmentation in OCT

Problem Statement
Accurate segmentation of the choroidal layer in Optical Coherence Tomography (OCT) images is critical for diagnosing ocular diseases such as age-related macular degeneration (AMD), central serous chorioretinopathy (CSCR), and diabetic retinopathy, as well as for assessing systemic conditions like cardiovascular disease. However, manual annotation of the choroid is time-consuming (taking 20–25 minutes per scan) and prone to error due to image artifacts. OCT images present specific challenges, including speckle noise, low contrast at the choroid-scleral interface, vessel shadowing, and anatomical variability. While deep learning models like U-Net and its variants have advanced medical image segmentation, they often struggle with the low-contrast boundaries and structural variability inherent to the choroid. Existing automated methods are frequently limited to handcrafted preprocessing, lack global context modeling, or rely on volumetric data that is not always available in standard 2D clinical workflows.

Methodology
The authors propose HAG-EUNet, a hybrid deep learning framework designed for automated choroid segmentation and clinical biomarker extraction. The methodology comprises three main components:

Robust Preprocessing Pipeline: A three-stage pipeline is employed to enhance structural clarity while preserving anatomical boundaries. This includes:
- Pre-smoothing: Gaussian filtering ( $\sigma = 0.4$ ) for initial noise reduction.
- Denoising: Bilateral filtering ( $d=5, \sigma_c=10, \sigma_s=10$ ) to reduce speckle noise while preserving edges.
- Enhancement: Adaptive Contrast Limited Adaptive Histogram Equalization (CLAHE) with conservative blending (60%) to improve local contrast without introducing artifacts.
- This pipeline reportedly yields a 3–4% improvement in Dice scores.
HAG-EUNet Architecture: The model is an encoder-decoder structure with specific enhancements:
- Encoder: Utilizes EfficientNet-B3 (pre-trained on ImageNet) to extract hierarchical features at five levels, balancing depth, width, and resolution for computational efficiency.
- Bottleneck: Integrates Atrous Spatial Pyramid Pooling (ASPP) with dilation rates of 1, 6, 12, and 18 to capture multi-scale contextual information, and a Transformer block with four-head self-attention to model global contextual relationships.
- Decoder: Features an attention-guided upsampling mechanism. It incorporates:
  - Dense Skip Connections: Aligning encoder features with upsampled decoder features via $1\times1$ convolutions.
  - Dual Attention: Squeeze-and-Excitation (SE) for channel attention and spatial attention maps derived from max/average pooling.
  - Residual Connections: To stabilize training and improve gradient flow.
- Feature Fusion: A Feature Pyramid Network (FPN) fuses features from all four decoder stages.
- Edge Refinement: A dedicated module uses Sobel operators to extract horizontal and vertical edge features, fusing them with decoder outputs to refine boundary delineation.
Training Strategy:
- Loss Function: A composite loss combining Dice Loss (weight 0.7) and Focal Loss (weight 0.3, $\alpha=0.25, \gamma=2.0$ ) to address class imbalance and boundary sensitivity.
- Deep Supervision: Auxiliary outputs at four decoder stages (7.5% weight each) are used alongside the main output (70% weight) to enhance gradient flow.
- Dataset: The model was trained on 385 OCT images, validated on 83, and tested on 83 images selected from a public dataset (C8). Ground truth masks were generated using DeepGPET and manually corrected by an expert ophthalmologist.

Key Contributions

Novel Architecture: The introduction of HAG-EUNet, which uniquely combines EfficientNet-B3, Transformer attention, ASPP, multi-form attention, and edge refinement specifically for choroid segmentation.
Optimized Preprocessing: A conservative three-stage pipeline that reduces noise while retaining anatomical integrity, demonstrating measurable performance gains.
Clinical Deployment: The development of a FastAPI-based system capable of real-time inference and the automated extraction of 32 clinically relevant biomarkers (including thickness metrics, geometric features, texture descriptors, and vascular characteristics), eliminating the manual annotation burden.

Results

Quantitative Performance: On the test set of 83 images, HAG-EUNet achieved a Dice score of 95.99% ± 1.86%, an IoU of 92.36%, and a Precision of 95.89%.
- It outperformed state-of-the-art baselines including Swin-UNet, ConvNeXt-UNet, U-Net, Attention U-Net, U-Mamba, U-Net++, DeepLabV3+, and TransUNet.
- Compared to the strongest baseline (TransUNet), it improved the Dice score by 0.72 percentage points.
- It demonstrated higher stability with a standard deviation of 1.86%, representing a ~29% reduction in variance compared to DeepLabV3+.
- The model is computationally efficient, using 37.5% fewer parameters (25M) than DeepLabV3+ (40M).
Ablation Studies: Removing individual components confirmed their necessity. The Transformer block removal caused the largest performance drop (−1.02% Dice), followed by ASPP (−0.95%) and Edge Refinement (−0.78%).
Biomarker Validation: Automated measurements for four primary biomarkers (mean thickness, subfoveal thickness, area, and maximum thickness) showed strong agreement with expert ground truth (Pearson $r$ ranging from 0.9288 to 0.9623; Spearman $\rho$ from 0.9436 to 0.9686) and low bias in Bland–Altman analysis.
Qualitative Analysis: Visual results indicate the model successfully delineates boundaries in low-contrast areas and presence of artifacts, though some challenges remain with abrupt thickness changes and highly curved boundaries.

Significance and Claims
The paper claims that HAG-EUNet provides a reliable, precise, and clinically deployable solution for choroid segmentation. By achieving high accuracy on standard 2D B-scans, the framework enables the automated extraction of 32 biomarkers to support responsible, fast, and reliable AI-based ophthalmic analysis. The authors position the work as a step toward automated quantitative OCT analysis, facilitating early disease detection, longitudinal monitoring, and cardiovascular risk assessment.

Limitations and Future Work
The authors modestly acknowledge several limitations:

Generalizability: The dataset consists of images from a single source and a single imaging device, which may limit applicability across different acquisition settings.
Dimensionality: The current framework processes 2D B-scans individually without modeling volumetric consistency across adjacent slices.
Annotation: Ground truth was provided by a single expert; multi-annotator validation would strengthen the evaluation.
Clinical Validation: While segmentation accuracy is high, prospective clinical studies are required to validate the direct relevance of extracted biomarkers to patient outcomes.

Future work is proposed to include multi-center evaluations, extension to full 3D volumetric analysis, integration of uncertainty estimation for explainability, and the segmentation of choroidal blood vessels to derive the Choroidal Vascularity Index (CVI).

Technical Summary: HAG-EUNet for Choroid Segmentation in OCT

More like this