Toponet: a topology-guided dual-branch framework for medical image classification

Volumetric medical image classification remains challenging due to limited robustness and interpretability, largely because modern models based on CNNs and Transformers primarily rely on appearance-driven cues and struggle to explicitly capture global anatomical topology, such as connectivity, loops, and cavities. To address this limitation, we propose TopoNet, a topology-guided dual-branch framework that integrates multiscale topological structures into volumetric representation learning. Specifically, TopoNet computes cubical persistent homology to obtain persistence diagrams and distills robust topological signals from background noise via statistical confidence bands, enabling adaptive and noise-robust filtering without relying on rigid thresholds. Crucially, to overcome the spatial information loss inherent in traditional vectorization methods, the selected persistence features are mapped back to the voxel domain. This process restores spatial context to generate spatially registered topological structures, forming an explicit topology modality aligned with the input images and providing spatially grounded structural cues that improve interpretability. The image and topology modalities are encoded in parallel and fused via stage-wise cross-branch attention to facilitate complementary feature interactions across scales. We validate TopoNet on two datasets, where it consistently outperforms strong baselines across key metrics, achieving \((0.8197 ± 0.0587)\) in Accuracy and \((0.8123 ± 0.1127)\) in AUC on the MSD Task07 Pancreas dataset, and \((0.9333 ± 0.0079)\) in Accuracy and \((0.9267 ± 0.0437)\) in AUC on the BraTS 2018 dataset. Overall, these results indicate that topology-guided multimodal fusion provides a promising avenue for improving performance, robustness, and clinical interpretability in volumetric medical image classification. The code is available at https://github.com/hbqdcq/TopoNet.