AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Topology-Reinforced Swin Transformer for Medical Image Analysis

Pengfei Gu, Huimin Li, Guangyu Meng, Hao Zheng, Haoteng Tang, Bin Fu, Danny Z Chen

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: Token merging and masked pretraining

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a topology-reinforced Swin Transformer that explicitly represents and incorporates multi-scale topological patterns (connected components, loops) into hierarchical ViT features for medical image analysis.

Problem

Hierarchical Vision Transformers such as Swin Transformer lack an explicit architectural scheme to represent and incorporate multi-scale topological structure (connected components, loops) in medical images, from microaneurysm-scale lesions to organ-level boundaries.

Approach

Adds an explicit topology-aware mechanism to the Swin Transformer hierarchy so that multi-scale topological features (connected components and loops) are represented and fed into the network; exact mechanism not stated in the abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Band-Aware H0/H1 residual refinement (ablation)Topo-SwinV2-B / Topo-Swin-UNet · CBIS-DDSM / Synapseaccuracy / Dice: 73.58 -> 75.62 acc; 81.03 -> 82.19 DiceImproved
inject topology as attention-logit bias rather than concatenation (ablation)TRiST · CBIS-DDSMaccuracy: 72.73 -> 75.62Improved
inject topology as attention bias rather than residual gating (ablation)TRiST · SynapseDice: 80.96 -> 82.19Improved
progressive band routing instead of uniform short band (ablation)TRiST · SynapseDice: 77.19 -> 82.19Improved
progressive band routing instead of uniform medium band (ablation)TRiST · CBIS-DDSMAUC: 82.19 -> 82.01Mixed
compute topology natively at all four stages instead of stage 1 only (ablation)TRiST · CBIS-DDSMaccuracy: 74.01 -> 75.62Improved
compute PH natively per stage instead of downsampling stage-1 topology (ablation)TRiST · SynapseHD95: 22.05 -> 11.10Improved
add topology representation to Swin-UNetSwin-UNet · SynapseDice: 79.76 -> 82.19Improved
add topology to SwinV2-BSwinV2-B · RetinaMNISTaccuracy: 67.67 -> 70.84Improved
3D methods versus topology-enhanced 2D modelTopo-Swin-UNet · SynapseDice: 82.19 vs 86.62 (UNETR++)Hurt
choice of topology descriptor (PI, BC, PS, PL) (ablation)Topo-SwinV2-B · classification datasetsaccuracyMixed

Experimental setup

Builds on: Swin Transformer

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy