💻 computer science

TopoNet: A Topology-Guided Dual-Branch Framework for Medical Image Classification

TopoNet is a topology-guided dual-branch framework that enhances volumetric medical image classification by integrating spatially registered topological features derived from persistent homology with appearance-based cues through cross-branch attention, thereby improving accuracy, robustness, and interpretability.

Original authors: Jin Sun, Nan Gao, Wenqi Tao, Zaifang Zhang, Chenghao Shao

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jin Sun, Nan Gao, Wenqi Tao, Zaifang Zhang, Chenghao Shao

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

Imagine you are trying to identify a specific type of cloud in the sky. Most computer programs today act like a painter who only looks at the colors and textures of the cloud. They say, "This cloud looks fluffy and white, so it must be a cumulus." But what if the shape of the cloud is actually a perfect donut, or a twisted loop? A painter might miss that, but a mathematician who studies topology (the science of shapes and connections) would instantly spot the hole in the middle.

This is exactly the problem doctors face with 3D medical scans. Current AI models are great at spotting "fluffy" textures, but they often miss the big picture: the loops, tunnels, and holes inside organs that signal disease. For example, a healthy pancreatic duct is like a smooth, continuous tube. A diseased one might be broken, fragmented, or twisted into a weird loop. Standard AI might get confused by the noise, but it needs to see the shape of the break to be sure.

Enter TopoNet, a new framework created by researchers at Shanghai University and their hospital partners. Think of TopoNet not just as a painter, but as a detective with two sets of eyes.

The Two-Branch Detective

TopoNet works like a dual-brain system:

The "Texture" Eye: This branch looks at the medical image just like a normal AI, studying the colors, shadows, and pixel details.
The "Shape" Eye: This is the magic part. It uses a mathematical tool called persistent homology to scan the image for invisible shapes—like connected blobs, loops, and empty cavities.

But here's the tricky part the paper solves: usually, when mathematicians find these shapes, they turn them into a boring list of numbers (like a spreadsheet). This is like taking a map of a city and turning it into a list of street names; you lose the sense of where the streets actually are. The paper argues that this "spatial detachment" is a major flaw. TopoNet fixes this by taking those abstract shapes and mapping them right back onto the 3D image, pixel by pixel. Now, the AI knows exactly where the loop or the hole is located, not just that it exists.

Filtering Out the Noise

Medical scans are often messy, full of "static" that looks like tiny, fake loops or holes. If you just grab every shape you see, you'll get confused. TopoNet uses a clever trick called confidence bands. Imagine you are trying to hear a friend's voice in a noisy room. Instead of shouting at everyone, you use a special filter that only lets through the voices that are loud and clear enough to be real. TopoNet does this mathematically: it separates the "real" disease shapes from the "fake" noise without needing a human to guess a cutoff point.

The Team-Up

Once TopoNet has its "Texture" view and its "Shape" view, it doesn't just slap them together. It uses a cross-attention mechanism. Think of this as a conversation between the two eyes. The "Shape" eye whispers to the "Texture" eye, "Hey, look right here! There's a loop in the duct that you missed!" This helps the AI focus on the most important structural clues, making it much harder to be tricked by confusing textures.

Did it Work?

The researchers tested this on two real-world datasets to see if it could tell the difference between healthy and diseased organs.

The Pancreas Test: They looked at CT scans of pancreas tumors. TopoNet achieved an Accuracy of 0.8197 ± 0.0587 and an AUC of 0.8123 ± 0.1127. This was better than many other strong AI models (like ResNet50 or ViT) that only looked at textures.
The Brain Tumor Test: They also tested it on brain MRI scans to distinguish between high-grade and low-grade gliomas. Here, TopoNet scored an Accuracy of 0.9333 ± 0.0079 and an AUC of 0.9267 ± 0.0437.

The paper explicitly rules out the idea that simply adding more data or using a "bigger" standard AI is the answer. They found that models relying only on texture often fail when the disease changes the organ's structure but not its color. They also showed that just listing the shapes (without mapping them back to the image) doesn't work as well as their dual-branch approach.

The Verdict

The authors suggest that by giving AI the ability to "see" the loops and holes in 3D space, we can make medical diagnosis more reliable and easier to understand. The results show that this "topology-guided" approach is a promising way to help computers understand the complex architecture of the human body, not just its surface appearance. While the paper notes that more testing on larger groups of patients is needed to be absolutely sure, the current experiments suggest that looking at the shape of the disease is a powerful new tool for doctors.

Technical Summary: TopoNet: A Topology-Guided Dual-Branch Framework for Medical Image Classification

Problem Statement

Volumetric medical image classification faces significant challenges regarding robustness and interpretability. Current state-of-the-art models, primarily based on Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), rely heavily on appearance-driven cues (texture and local intensity). These models often struggle to explicitly capture global anatomical topology, such as connectivity, loops, and cavities, which are critical for clinical diagnosis. For instance, in pancreatic ductal adenocarcinoma (PDAC), the distinction between healthy and diseased tissue often relies on topological changes like ductal fragmentation or abnormal dilation, which local pixel-level architectures fail to guarantee.

Furthermore, existing attempts to integrate Topological Data Analysis (TDA) into deep learning suffer from two main limitations:

Sensitivity to Noise: Traditional methods often use rigid, manually tuned thresholds to filter topological features, which can discard subtle but clinically relevant signals or retain noise due to case-to-case variability.
The "Spatial Detachment" Problem: Most approaches vectorize topological features (e.g., into persistence images or Betti curves) and concatenate them with neural features. This process discards the spatial coordinates of the features, preventing the network from localizing where specific topological anomalies occur, thereby creating a representational mismatch between spatially-aware image features and spatially-agnostic topological vectors.

Methodology: TopoNet

To address these limitations, the authors propose TopoNet, a topology-guided dual-branch framework that integrates multiscale topological structures directly into volumetric representation learning while preserving spatial context.

1. Statistical Topology Extraction via Confidence Bands

Instead of using fixed thresholds, TopoNet employs a statistically grounded strategy to distinguish significant topological signals from noise:

Cubical Persistent Homology: The method computes persistent homology on cubical complexes derived from the voxel grid, tracking the evolution of $H_0$ (connected components), $H_1$ (loops), and $H_2$ (cavities) across sublevel set filtrations.
Confidence Band Filtering: To handle topological noise, the framework uses a subsampling-based approach to construct $(1-\alpha)$ -level confidence bands for the persistence diagram. Features with persistence values ( $d-b$ ) falling within the confidence band (defined by the bottleneck distance $\epsilon_\alpha$ ) are treated as noise. Only features satisfying $d-b > 2\epsilon_\alpha$ are retained as statistically significant. This eliminates the need for rigid, manual threshold tuning.

2. Dimension-Aware Topology-to-Voxel Mapping

A core innovation of TopoNet is the restoration of spatial context to topological features, solving the "spatial detachment" problem:

Localization: Selected persistence pairs are mapped back to the voxel domain.
- $H_0$ : Connected component masks are generated by identifying seed voxels and extracting components at specific pre-merge thresholds.
- $H_1$ : Representative 1-cycles (loops) are reconstructed via matrix reduction and rasterized into binary loop masks, optionally dilated to form tubular neighborhoods.
- $H_2$ : Cavity shells are identified using "destroyer" seeds and constructed via geodesic region growing and morphological operators.
Result: This process generates explicit, spatially registered binary topology masks ( $T$ ) that are strictly aligned with the input image volume ( $I$ ), creating a complementary "topology modality."

3. Dual-Branch Architecture with Cross-Modal Fusion

TopoNet utilizes a dual-branch 3D architecture (based on DenseNet121) to process the image and topology modalities in parallel:

Mono-modal Feature Extraction: Two independent branches extract multi-scale features. The topology branch utilizes Group Normalization (GN) instead of Batch Normalization to handle the high sparsity of topology-derived masks.
Cross-Branch Interaction Module (CBIM): At intermediate stages (Stage 3 and 4), the branches interact via a specialized fusion mechanism:
- Texture Enhancement Module: A lightweight depth-wise convolution module enhances local texture features in the image branch.
- Topology Enhancement Module: A sparsity-aware module refines topology features, normalizing by valid voxel counts to prevent background zeros from contaminating the representation.
- Cross-Attention: Image features act as queries, while topology features provide keys and values. This allows the network to condition image representations on topological guidance.
- Gated Residual Fusion: The topology-guided features are injected into the image branch via a learnable gating mechanism, allowing the model to adaptively control the strength of the topological prior.
Classification Head: The final fused features undergo Attention Pooling, which adaptively aggregates voxel-wise features based on learned attention weights, suppressing background noise before classification.

Key Contributions

The paper outlines four primary contributions:

Statistically Grounded Extraction: A framework using confidence bands to identify significant persistent homology features on cubical complexes, separating robust clinical signals from noise without rigid thresholds.
Spatial Restoration: A dimension-aware topology-to-voxel mapping pipeline that converts abstract topological summaries ( $H_0/H_1/H_2$ ) into spatially aligned, voxel-wise binary masks, bridging the gap between TDA and grid-based deep learning.
TopoNet Architecture: A dual-branch network that encodes image and topology modalities in parallel, utilizing cross-branch attention to guide the network toward structurally significant regions.
Empirical Validation: Extensive experiments demonstrating consistent improvements over strong baselines on two distinct datasets.

Experimental Results

The authors evaluated TopoNet on two datasets using 5-fold cross-validation:

MSD Task07 Pancreas Dataset (PDAC vs. Non-PDAC):
- TopoNet achieved an Accuracy of 0.8197 ± 0.0587 and an AUC of 0.8123 ± 0.1127.
- It outperformed strong CNN (ResNet50, DenseNet121, EfficientNet) and Transformer (ViT, Swin) baselines.
- Notably, TopoNet showed significantly lower variance in Accuracy and F1-score compared to baselines, indicating greater stability.
BraTS 2018 Dataset (Glioma Grading: HGG vs. LGG):
- TopoNet achieved an Accuracy of 0.9333 ± 0.0079 and an AUC of 0.9267 ± 0.0437.
- It demonstrated superior generalization to MRI modalities compared to baselines, with highly stable predictions across folds.

Ablation Studies confirmed that:

Confidence-band selection outperforms fixed-threshold strategies.
Combining all homology dimensions ( $H_0 + H_1 + H_2$ ) yields the best performance.
The "No Inversion" filtration strategy (sublevel set on raw intensity) was optimal for pancreatic CT, highlighting the importance of task-specific filtration semantics.
The specific direction of cross-modal interaction (Topology-guided Image enhancement, $T \to I$ ) was more effective than bidirectional or sequential alternatives.
Removing any component (Texture enhancement, Topology enhancement, Cross-attention, or Attention pooling) degraded performance, validating the necessity of the full architecture.

Visualization: Grad-CAM visualizations showed that TopoNet focuses on disease-relevant structural patterns (e.g., ductal trajectories) aligned with topological features, whereas baselines often produced diffuse, texture-driven activations.

Significance and Claims

The paper claims that TopoNet provides a promising avenue for improving the performance, robustness, and clinical interpretability of volumetric medical image classification. By fundamentally integrating explicit structural cues into the deep learning pipeline rather than treating topology as a global statistic, the framework addresses the "spatial detachment" inherent in previous TDA methods.

The authors emphasize that their approach:

Enhances Interpretability: By mapping topological features back to the voxel domain, the model provides spatially grounded structural cues that align with clinical reasoning (e.g., ductal connectivity).
Improves Robustness: The confidence-band filtering and gated fusion mechanisms allow the model to adapt to noise and anatomical variability, reducing reliance on incidental textures.
Offers a Plug-and-Play Paradigm: The topology-guided learning framework is shown to be effective across various backbone architectures (CNNs and Transformers), suggesting broad applicability.

The authors conclude that while the method shows promise, future work is needed to model synergistic interactions between homology dimensions (e.g., using Graph Neural Networks) and to validate the approach on larger, multi-center cohorts to assess generalizability across diverse scanner protocols.