💻 computer science

TFFM: Topology-Aware Feature Fusion Module via Latent Graph Reasoning for Retinal Vessel Segmentation

This paper introduces TFFM, a topology-aware framework that combines a latent graph-based feature fusion module with a hybrid loss function to achieve state-of-the-art retinal vessel segmentation with significantly improved vascular connectivity and reduced fragmentation.

Original authors: Iftekhar Ahmed, Shakib Absar, Aftar Ahmad Sami, Shadman Sakib, Debojyoti Biswas, Seraj Al Mahmud Mostafa

Published 2026-02-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Iftekhar Ahmed, Shakib Absar, Aftar Ahmad Sami, Shadman Sakib, Debojyoti Biswas, Seraj Al Mahmud Mostafa

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

The Big Problem: The "Shattered Glass" Effect

Imagine you are looking at a map of a city's road network. Your goal is to trace every single street. A standard computer program (an AI) might do a great job of coloring in the pixels where the roads are. It gets 95% of the dots right.

However, there's a catch: the AI might draw the roads as a bunch of disconnected, floating dots. It misses the tiny bridges connecting them. To a human, the road looks broken. To a doctor trying to analyze blood flow, this is useless because a broken road isn't a road at all—it's just a pile of gravel.

In the world of eye scans (retinal images), this is called the "shattered vessel" phenomenon. The AI gets the pixels right but breaks the topology (the connectedness) of the blood vessels.

The Solution: TFFM (The "Traffic Controller")

The authors of this paper built a new system called TFFM (Topology-Aware Feature Fusion Module). Think of it as upgrading the AI from a simple "dot-painter" to a "traffic controller."

Here is how it works, step-by-step:

1. The "Latent Graph" (The Invisible Web)

Standard AI looks at an image in small, local chunks (like looking at a map through a tiny straw). It sees a road here and a road there, but it doesn't know they are connected.

The TFFM takes those local chunks and projects them into a "latent graph space."

Analogy: Imagine taking all the scattered road pieces and laying them out on a giant table. The TFFM then stretches invisible elastic strings between pieces that should be connected based on their shape and direction, even if they are far apart on the image. It creates a "web" of connections that the AI can see and reason about.

2. Graph Attention Networks (The Smart Detective)

Once the AI has this web, it uses a tool called Graph Attention Networks (GAT).

Analogy: Instead of just looking at one street corner, the AI acts like a detective who can see the whole city block at once. It asks, "If I follow this road, does it logically lead to that other road?" It uses this global view to fix the broken bridges that the standard AI missed.

3. The "Hybrid Scorecard" (The Double-Check System)

To teach the AI to do this, the authors created a special grading system (Loss Function) with two parts:

Part A (Tversky Loss): This checks if the AI found the right amount of road. It makes sure the AI doesn't ignore the thin, hard-to-see veins (which are often missed).
Part B (Soft clDice Loss): This is the "connectivity coach." It specifically penalizes the AI if it breaks a road.
- Analogy: Imagine a teacher grading a drawing. Part A gives points for coloring the lines red. Part B gives extra points for making sure the lines actually touch at the intersections. If the lines are floating in the air, Part B deducts points, forcing the student to fix the breaks.

The Results: A Connected Map

The team tested this new system on a dataset of eye images called Fundus-AVSeg.

The Score: They achieved a 90.97% accuracy score (Dice score), which is the best result reported so far.
The Real Win: The most important result wasn't just the score, but the connectivity.
- Standard AI models broke the blood vessels into many tiny, disconnected pieces (fragmentation).
- The TFFM system reduced these broken pieces by about 38%.
- Instead of a shattered map, the AI produced a continuous, unbroken tree of blood vessels that looks like a real biological network.

Why This Matters (According to the Paper)

The paper emphasizes that for doctors to use these scans for automated analysis (like measuring how twisted a vessel is), the map must be whole. You cannot measure the length of a road if the road is in three separate pieces.

By using this "traffic controller" (TFFM) and the "connectivity coach" (the special scoring system), the AI finally produces blood vessel maps that are not just accurate in color, but structurally sound, making them ready for the next step of automated medical analysis.

In short: The paper teaches the AI to stop just "coloring dots" and start "drawing connected roads," ensuring the final map of the eye's blood vessels is whole, continuous, and usable.

Technical Summary: TFFM for Retinal Vessel Segmentation

1. Problem Statement

Retinal artery and vein segmentation is critical for diagnosing systemic cardiovascular conditions, hypertension, and diabetic retinopathy. However, a significant semantic gap exists between standard deep learning models and clinical requirements. While conventional Convolutional Neural Networks (CNNs) often achieve high pixel-level accuracy (e.g., Dice scores), they frequently produce topologically disjointed segmentations. These outputs contain gaps, discontinuities, and fragmented vessel trees, rendering them unsuitable for graph-based clinical analysis, such as automated biomarker quantification or tortuosity measurement. Standard metrics like Dice and Intersection over Union (IoU) fail to penalize these topological errors, masking the "shattered vessel" phenomenon where a model achieves high accuracy but produces clinically unusable, fragmented structures.

2. Methodology

The authors propose TFFM (Topology-Aware Feature Fusion Module), a framework designed to maintain vascular connectivity by integrating topological reasoning directly into the feature extraction and fusion process, rather than relying solely on post-hoc loss functions.

2.1. Baseline Architecture

The foundation of the framework is a U-Net++ backbone enhanced with Attention Gates, selected through ablation studies for its superior ability to capture fine-grained morphological details compared to standard U-Net, SwinUNETR, and TransUNet. The encoder path utilizes EfficientNet-B0, which was found to outperform larger variants (like EfficientNet-B4) and deeper networks (ResNet-101) in preserving fine vessel details while maintaining computational efficiency.

2.2. Topological Feature Fusion Module (TFFM)

The core innovation is the TFFM, which projects local feature maps into a latent graph space to reason about global connectivity. The module operates in four stages:

Scale-Adaptive Abstraction: Features are compressed and pooled into grids of varying sizes ( $G_l$ ) at different decoder levels.
Dynamic Graph Construction: A feature-space graph is constructed where nodes represent flattened feature vectors. Edges are defined via a sparsified cosine similarity matrix, selecting the top- $k$ neighbors to form an adjacency structure.
Graph Attention Reasoning: A Graph Attention Network (GAT) processes these nodes. It computes masked attention scores to aggregate information from non-local, topologically relevant regions, effectively capturing global structural dependencies missed by fixed receptive fields.
Adaptive Integration: The graph-derived features are refined using Channel and Spatial Attention Modules and a vesselness gating mechanism. A learnable gated residual connection blends these global topological cues with the original local features, allowing the network to selectively inject connectivity information without destroying fine local structures.

2.3. Hybrid Objective Function

To address both class imbalance and topological continuity, the model employs a composite loss function:

Tversky Loss: Used to mitigate the extreme foreground-background imbalance by asymmetrically weighting false positives and negatives. The authors tune parameters ( $\alpha=0.65, \beta=0.35$ ) to prioritize recall, as missed vessel segments (false negatives) are more detrimental to topology than background noise.
Soft clDice Loss: A differentiable topological regularizer that operates on continuous probability maps. It computes the harmonic mean of Topological Precision and Sensitivity based on soft-skeletonized predictions, explicitly penalizing breaks in the vessel skeleton.
Total Loss: $L_{total} = L_{Tv} + 0.5 L_{clDice}$ .

3. Key Contributions

Architecture-Integrated Topology: Unlike previous works that apply topological constraints only during optimization (via loss functions), this work integrates topological reasoning directly into the feature hierarchy via the TFFM and GATs.
Latent Graph Reasoning: The introduction of a module that maps local features to a latent graph space allows the network to model global vascular connectivity beyond the local receptive field of standard convolutions.
Systematic Ablation: The study provides a rigorous selection process for the backbone (U-Net++ w/ Attention Gates), encoder (EfficientNet-B0), and loss formulation, demonstrating that architectural choices significantly impact topological fidelity.
Zero-Shot Generalization: The framework demonstrates robustness across diverse external datasets without fine-tuning, suggesting the learned topological constraints are generalizable across imaging protocols.

4. Experimental Results

The method was evaluated on the Fundus-AVSeg dataset and five external benchmarks (DRIVE, CHASEDB1, HRF, RETA, STARE).

Performance on Fundus-AVSeg:
- Achieved a combined Dice score of 90.97% and a 95% Hausdorff Distance (HD95) of 3.50 pixels.
- Topological Improvement: The method reduced vessel fragmentation by approximately 38% relative to baselines (from 44.7 to 27.7 predicted components per image).
- Betti Error: The Betti number error (measuring connected component differences) dropped significantly from 43.0 to 26.0 upon adding TFFM.
- Junction Detection: Junction F1-scores improved to 65.44%, though recall for bifurcation points remains a challenge (0.44 for arteries, 0.52 for veins).
Cross-Dataset Generalization:
- In zero-shot inference on external datasets, the model maintained Dice scores between 79.94% and 82.18%.
- Topological consistency (clDice) remained high (68.71%–73.57%), significantly outperforming traditional methods in preserving vascular connectivity without domain adaptation.

5. Significance and Claims

The paper claims that the primary significance of TFFM lies in bridging the gap between pixel-level accuracy and clinical utility. By ensuring topological coherence, the framework produces vascular trees that are viable for automated graph-based analysis, such as biomarker quantification.

The authors emphasize that while standard metrics (Dice, IoU) can be misleading by rewarding fragmented predictions, their hybrid approach successfully balances region overlap with structural integrity. They note that while the method significantly reduces fragmentation and improves global connectivity, bifurcation completeness remains a limitation, with junction recall rates suggesting that extreme local complexity at crossing points still requires specialized attention mechanisms. The work concludes that their framework generates topologically robust segmentations suitable for downstream clinical tasks, offering a path toward more reliable automated retinal analysis.