TopoCL: Topological Contrastive Learning for Medical Imaging

Guangyu Meng; Pengfei Gu; Peixian Liang; John P. Lalor; Erin Wolf Chambers; Danny Z. Chen

cs.CVMar 15, 2026

TopoCL: Topological Contrastive Learning for Medical Imaging

Authors: Guangyu Meng, Pengfei Gu, Peixian Liang, John P. Lalor, Erin Wolf Chambers, Danny Z. Chen

Organizations: Dept. of Computer Science and Engineering, University of Notre Dame · The University of Texas Rio Grande Valley · Dept. of IT, Analytics, and Operations, University of Notre Dame

Abstract

Contrastive learning (CL) has become a powerful approach for learning representations from unlabeled images. However, existing CL methods focus predominantly on visual appearance features while neglecting topological characteristics (e.g., connectivity patterns, boundary configurations, cavity formations) that provide valuable cues for medical image analysis. To address this limitation, we propose a new topological CL framework (TopoCL) that explicitly exploits topological structures during contrastive learning for medical imaging. Specifically, we first introduce topology-aware augmentations that control topological perturbations using a relative bottleneck distance between persistence diagrams, preserving medically relevant topological properties while enabling controlled structural variations. We then design a Hierarchical Topology Encoder that captures topological features through self-attention and cross-attention mechanisms. Finally, we develop an adaptive mixture-of-experts (MoE) module to dynamically integrate visual and topological representations. TopoCL can be seamlessly integrated with existing CL methods. We evaluate TopoCL on five representative CL methods (SimCLR, MoCo-v3, BYOL, DINO, and Barlow Twins) and five diverse medical image classification datasets. The experimental results show that TopoCL achieves consistent improvements: an average gain of +3.26% in linear probe classification accuracy with strong statistical significance, verifying its effectiveness.

Figures & tables

Figure 1: Visualization of failure cases corrected by TopoCL on ISIC2019 [ 39 ] . Two dermatofibroma (DF) cases are misclassified by the baseline MoCo-v3 [ 9 ] as actinic keratosis (AK, top) and melanocytic nevi (NV, bottom). MoCo-v3+TopoCL correctly classifies both as DF by capturing characteristic topological features: circular-to-oval boundary patterns with uniform internal connectivity (top) and radial pigmentation structures with a consistent boundary configuration despite hair overlay (bottom). Grad-CAM [ 35 ] heatmaps show that TopoCL focuses on lesion boundaries and internal structural patterns, while the baseline exhibits scattered attention on peripheral or irrelevant regions.

Figure 2: An overview of our TopoCL framework. (a) Topological Contrastive Learning : Given an input image $x$ , we generate two views with topology-weak ( $x_{\text{topo}}^{w}$ ) and topology-strong ( $x_{\text{topo}}^{s}$ ) augmentations, and compute their persistence diagrams $\text{topo}^{w}$ and $\text{topo}^{s}$ . The visual encoder processes $x_{\text{topo}}^{w}$ , $x_{\text{topo}}^{s}$ while the topology encoder processes $\text{topo}^{w}$ , $\text{topo}^{s}$ , with projection heads producing visual features $\mathbf{f}^{w}$ , $\mathbf{f}^{s}$ and topological features $\mathbf{t}^{w}$ , $\mathbf{t}^{s}$ . The TopoCL MoE module fuses these features through five experts (the colors distinguish the fusion strategies: Vis-Only, Topo-Only, Concat, Gated, and Cross-Attn) with learned gating weights, producing fused representations $\mathbf{h}^{w}$ , $\mathbf{h}^{s}$ . A projection head maps these to the final representations $\mathbf{z}^{w}$ , $\mathbf{z}^{s}$ , optimized via contrastive loss. (b) The structure of H-Topo. Encoder : It uses hierarchical self- and cross-attention to capture topological features from PDs with both $H_{0}$ and $H_{1}$ homology dimensions.

Category	Operations	Topological Effect
Homeomorphism	Flips, rotations	Topology-preserving ( $d_{B}^{\text{rel}}\!\approx\!0$ )
Boundary perturb.	Gaussian noise	Perturb boundary structures
Smoothing	Gaussian blur	Merge/separate features
Intensity	Contrast, brightness	Shift feature prominence
Morphological	Dilation, erosion	Modify connectivity and holes

Table 1: Topology-aware augmentation categories. All topological effects (except homeomorphism) are parameter-dependent.

Dataset	Modality	Classes	Train / Test
PathMNIST	Colon Pathology	9	89,996 / 7,180
OCTMNIST	Retinal OCT	4	97,477 / 1,000
OrganSMNIST	Abdominal CT (Sagittal)	11	13,940 / 8,829
ISIC2019	Skin Lesion Dermoscopy	8	25,331 / 8,238
Kvasir	GI Endoscopy	8	3,200 / 800

Table 2: Dataset statistics for the five medical image datasets.

	Path		OrganS		OCT		ISIC		Kvasir		Avg.
Method	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC
SimCLR [ 8 ]	$92.57_{\pm 0.31}$	$99.12_{\pm 0.11}$	$77.12_{\pm 0.26}$	$96.41_{\pm 0.26}$	$69.41_{\pm 0.56}$	$91.83_{\pm 0.43}$	$66.04_{\pm 2.29}$	$88.01_{\pm 1.11}$	$74.33_{\pm 1.28}$	$95.89_{\pm 0.76}$	$75.89_{\pm 0.51}$	$94.25_{\pm 0.11}$
+TopoCL	$93.21_{\pm 0.14}^{*}$	$99.28_{\pm 0.05}$	$78.62_{\pm 0.06}^{***}$	$96.75_{\pm 0.33}^{*}$	$73.10_{\pm 0.21}^{***}$	$93.41_{\pm 0.26}^{**}$	$71.62_{\pm 0.21}^{**}$	$89.82_{\pm 0.77}^{**}$	$78.61_{\pm 0.41}^{**}$	$97.11_{\pm 0.32}^{*}$	$79.03_{\pm 0.10}^{***}$	$95.27_{\pm 0.15}^{***}$
MoCo-v3 [ 9 ]	$93.13_{\pm 0.42}$	$99.07_{\pm 0.09}$	$78.02_{\pm 0.34}$	$99.57_{\pm 0.31}$	$80.02_{\pm 0.34}$	$96.01_{\pm 0.04}$	$74.98_{\pm 2.01}$	$92.86_{\pm 1.12}$	$88.42_{\pm 2.01}$	$97.37_{\pm 2.01}$	$82.91_{\pm 0.66}$	$96.98_{\pm 0.60}$
+TopoCL	$94.55_{\pm 0.20}^{**}$	$99.24_{\pm 0.09}^{**}$	$80.58_{\pm 0.09}^{***}$	$98.75_{\pm 0.21}^{\dagger}$	$82.09_{\pm 0.26}^{***}$	$97.09_{\pm 0.13}^{***}$	$78.44_{\pm 0.83}^{*}$	$93.26_{\pm 0.97}$	$91.17_{\pm 0.83}^{*}$	$98.81_{\pm 0.46}$	$85.37_{\pm 0.11}^{***}$	$97.43_{\pm 0.27}$
BYOL [ 17 ]	$93.42_{\pm 0.09}$	$99.31_{\pm 0.14}$	$76.41_{\pm 0.19}$	$97.33_{\pm 0.08}$	$76.02_{\pm 0.26}$	$95.89_{\pm 0.25}$	$73.82_{\pm 3.19}$	$90.75_{\pm 0.79}$	$85.67_{\pm 0.98}$	$96.91_{\pm 0.13}$	$81.07_{\pm 0.49}$	$96.04_{\pm 0.20}$

Table 3: Performance comparison across five medical image datasets. The results show mean

\pm

standard deviation over five independent runs. Statistical significance is assessed via paired t-tests: ∗∗∗ :

p<0.001

, ∗∗ :

p<0.01

, ∗ :

p<0.05

. † indicates performance decrease. The Avg. columns report dataset-averaged metric values.

Configuration	OrganS	ISIC	Kvasir
TopoCL (weak+strong)	75.51 ±1.56	72.92 ±2.32	84.63 ±1.85
TopoCL (full image)	76.34 ±1.46	74.35 ±1.58	87.41 ±1.88
TopoCL (topo-weak+topo-weak)	78.98 ±0.37	76.39 ±0.55	90.01 ±0.28
TopoCL (topo-strong+topo-strong)	78.02 ±0.68	75.99 ±1.02	89.64 ±0.91
TopoCL (topo-weak+topo-strong)	80.58 ±0.99	78.44 ±0.83	91.17 ±0.43

Table 4: Ablation of topology-aware augmentation configurations using MoCo-v3+TopoCL. ACC results (%) are reported.

Configuration	Path	OrganS	OCT
w/o pretraining	$68.98_{\pm 1.43}$	$50.61_{\pm 0.87}$	$48.72_{\pm 1.77}$
w/o vis. pretrain	$81.53_{\pm 0.98}$	$64.72_{\pm 1.21}$	$66.98_{\pm 0.83}$
w/o topo. pretrain	$87.64_{\pm 0.92}$	$72.56_{\pm 1.11}$	$69.37_{\pm 0.89}$
w/o hier. attn	$86.51_{\pm 0.93}$	$72.85_{\pm 0.76}$	$67.62_{\pm 1.03}$
w/o self-attn	$91.24_{\pm 0.85}$	$76.03_{\pm 0.56}$	$70.28_{\pm 0.66}$
w/o cross-attn	$87.59_{\pm 0.85}$	$74.66_{\pm 0.71}$	$69.38_{\pm 0.57}$

Table 5: Ablation with SimCLR+TopoCL on three datasets. We evaluate: (1) pretraining strategies (rows 1-3); (2) hierarchical topology encoder components (rows 4-6); (3) MoE expert choices (rows 7-11). ACCs (%) averaged over five runs are reported.

Figure 3: Expert gating analysis for TopoCL integrated with BYOL on five datasets. Error bars show standard deviation across test samples. Five experts: visual-only (Vis), topology-only (Topo), concatenation (Conc), gated blending (Gate), and cross-attention (Attn).

Method	Train (h)	Params (M)	FLOPs (G)	Tput.	Overhead
SimCLR	1.58	24.89	4.11	668	-
+TopoCL	1.84	31.08	6.11	574	+16%
BYOL	1.72	53.28	2.88	614	-
+TopoCL	1.96	60.48	4.58	539	+14%
DINO	1.79	45.16	5.23	590	-
+TopoCL	1.99	51.00	7.03	531	+11%

Table 6: Computational cost comparison on ISIC2019.

Component	Training		Inference
Component	Time (h)	%	Params (M)	FLOPs (G)
Visual Encoder	1.06	57.61	24.03	4.11
Topo Encoder	0.35	19.02	3.83	1.99
MoE+Proj	0.43	23.37	3.23	0.01
Total	1.84	100.0	31.08	6.11

Table 7: Training and inference cost breakdown for SimCLR+TopoCL on ISIC2019.

Explore similar work

Jun 29, 2026

TopoAgent: An Agentic Framework for Automated Topology Learning in Medical Imaging

Topological data analysis (TDA), particularly persistent homology (PH), captures geometric structural properties in medical images (e.g., connected components, loops, shape characteristics), which conventional pixel-level deep learning approaches often neglect. While many topological descriptors are known for converting persistence diagrams (PDs) or raw images into topological feature vectors, existing methods mostly default to a single fixed descriptor (e.g., persistence images), leaving the diversity of topological representations largely unexplored. To the best of our knowledge, there is no known large language model (LLM)-based agentic framework that can automatically determine the most suitable topological descriptors for a given image dataset and produce the corresponding topological feature vectors for downstream tasks. To fill this gap, we propose TopoAgent, an LLM-based agentic framework that automates topology learning for medical image analysis.TopoAgent operates through a Perception--Reasoning--Action--Reflection loop supported by 21 domain-specific tools and dual memory that accumulates experience across runs. Its skill set is distilled from systematic evaluation of 15 topological descriptors across 26 datasets with six classifiers. TopoAgent analyzes input images and their topological characteristics, reasons about which topological descriptors best suit the input, and determines the optimal descriptor and its configuration, all without task-specific training.

Guangyu Meng, Pengfei Gu, Xueyang Li +3

Dept. of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN, USA · Dept. of Computer Science, The University of Texas Rio Grande Valley, Edinburg, TX, USA

Medical Image Analysis Persistent Homology Automated Feature Engineering Topological Representation Learning Topological Data Analysis

Ask PDF

Aug 4, 2026

Recurrent Contrastive Learning for Imbalanced Medical Image Classification

Medical image classification often suffers from class imbalance due to the inherent disparities in disease incidence. Existing approaches, such as class resampling and loss reweighting, mainly improve learning within the observed feature distribution, but do not explicitly enlarge the latent support region of tail classes. As a result, tail-class representations remain overly compact and are easily encroached upon by head classes, leading to biased decision boundaries. In this work, we propose Recurrent Contrastive Learning (RCL) for imbalanced medical image classification. RCL progressively expands the support region of tail classes by recurrently reusing historical feature states across training phases. Specifically, we adopt DINOv3 with LoRA adapters as the backbone to provide robust feature embeddings. We then devise a Temporal Memory Queue (TMQ) to preserve corpus-level features across training phases and provide diversified global references for contrastive learning. Based on TMQ, we construct Temporal Anchors (TARs) to form an anchor field around tail classes. This field enlarges the support region of tail classes, suppresses head-class encroachment, and improves inter-class separation. Extensive experiments on three imbalanced medical datasets demonstrate that RCL achieves consistent improvements over strong baselines. The code is available at https://github.com/dndins/RCL.

Zhiyuan Zhu, Xinling Meng, Junxuan Yu +15

School of Biomedical Engineering, Medical School, Shenzhen University, China · Ultra-X AI Lab, Shenzhen University, Guangdong, China · School of Biomedical Engineering, South-Central Minzu University, Wuhan, China +4

Class-Imbalanced Learning Contrastive Learning Medical Image Classification

Ask PDF

Mar 24, 2026

Bi-CRCL: Bidirectional Conservative-Radical Complementary Learning with Pre-trained Foundation Models for Class-incremental Medical Image Analysis

Class-incremental learning (CIL) in medical image-guided diagnosis requires retaining prior diagnostic knowledge while adapting to newly emerging disease categories, which is critical for scalable clinical deployment. This problem is particularly challenging due to heterogeneous data and privacy constraints that prevent memory replay. Although pretrained foundation models (PFMs) have advanced general-domain CIL, their potential in medical imaging remains underexplored, where domain-specific adaptation is essential yet difficult due to anatomical complexity and inter-institutional heterogeneity. To address this gap, we conduct a systematic benchmark of recent PFM-based CIL methods and propose Bidirectional Conservative-Radical Complementary Learning (Bi-CRCL), a dual-learner framework inspired by complementary learning systems. Bi-CRCL integrates a conservative learner that preserves prior knowledge through stability-oriented updates and a radical learner that rapidly adapts to new categories via plasticity-oriented learning. A bidirectional interaction mechanism enables forward transfer and backward consolidation, allowing continual integration of new knowledge while mitigating catastrophic forgetting. During inference, outputs from both learners are adaptively fused for robust predictions. Experiments on five medical imaging datasets demonstrate consistent improvements over state-of-the-art methods under diverse settings, including cross-dataset shifts and varying task configurations.

Xinyao Wu, Zhe Xu, Cheng Chen +3

Department of Biomedical Engineering, The Chinese University of Hong Kong, Hong Kong, China. · Department of Radiation Oncology, Columbia University Irving Medical Center and Data Science Institute, Columbia University, New York, NY, USA. · Department of Electrical and Electronic Engineering and School of Biomedical Engineering, The University of Hong Kong, Hong Kong, China. +2

Class-Incremental Learning Medical Image Analysis Medical Image Classification Catastrophic Forgetting Medical Imaging Foundation Models

Ask PDF