Mammo-Clustering: Context Clustering based Multi-view Tri Level Information Fusion for Lesion Location and Classification in Mammography

Shilong Yang, Chulong Zhang, Xiaokun Liang, Qi Zang, Juan Yu, Liang Zeng, Xiao Luo, Yadi Xing, Tingting Peng, Qi Li, Linlin Shen, Yaoqin XiearXiv 2025 · DOI · arXiv

Abstract

Breast cancer is a significant global health issue, and the diagnosis of breast cancer through imaging remains challenging. Mammography images are characterized by extremely high resolution, while lesions often occupy only a small portion of the image. Down-sampling in neural networks can easily lead to the loss of microcalcifications or subtle structures. To tackle these challenges, we propose a Context Clustering based triple information fusion framework. First, in comparison to CNNs or transformers, we observe that Context clustering methods are (1) more computationally efficient and (2) better at associating structural or pathological features. This makes them particularly well-suited for mammography in clinical settings. Next, we propose a triple information fusion mechanism that integrates global, feature-based local, and patch-based local information. The proposed approach is rigorously evaluated on two public datasets, Vindr-Mammo and CBIS-DDSM, using five independent data splits to ensure statistical robustness. Our method achieves an AUC of 0.828 ± 0.020 on Vindr-Mammo and 0.805 ± 0.020 on CBIS-DDSM, outperforming the second best method by 3.5% and 2.5%, respectively. These improvements are statistically significant (p<0.05), highlighting the advantages of the Context Clustering Network with triple information fusion. Overall, our Context Clustering framework demonstrates strong potential as a scalable and cost-effective solution for large-scale mammography screening, enabling more efficient and accurate breast cancer detection.

1. Introduction

Detecting breast cancer via mammography is a classic "needle in a haystack" problem. Mammograms possess extremely high resolutions (e.g., 3518 × 2800 pixels), yet critical lesions often occupy as little as a 10 × 10 pixel area. Traditional deep learning paradigms face significant hurdles in this domain. Convolutional Neural Networks (CNNs) rely on aggressive down-sampling that can erase fine microcalcifications, while Vision Transformers (ViTs) are computationally prohibitive at such high resolutions and often struggle with background noise. Furthermore, existing fusion methods often fail to deeply integrate global context with local high-resolution details.

2. Methodology

To address these limitations, we introduce Mammo-Clustering, an architecture that replaces rigid grids with Context Clustering (CoC) and employs a Tri-level Information Fusion Framework (TIFF).

2.1 Context Clustering (CoC)

Instead of standard convolutions, we treat images as unordered sets of 5D points (RGB values + xy coordinates). The network groups these points into clusters based on similarity. This approach preserves structural details without the information loss associated with aggressive down-sampling, aligning well with the clustered nature of pathological calcifications.

2.2 Tri-level Information Fusion Framework (TIFF)

The model extracts and fuses three distinct levels of information:

  • Global Information: A macroscopic feature map of the entire image.
  • Feature-based Local Information: Features extracted directly from the global feature map at Regions of Interest (RoIs) identified by a Saliency Map.
  • Patch-based Local Information: Detailed features from raw high-resolution image crops at the RoIs, processed through a secondary Local CoC network.
Detailed architecture of the Mammo-Clustering framework, illustrating the multi-view input, Context Clustering blocks, and the Tri-level Information Fusion mechanism.

2.3 Multi-view Integration

Following clinical practice, the model integrates features from four standard mammographic views (LCC, LMLO, RCC, RMLO) using an attention mechanism to produce the final classification.

3. Experiments and Results

The framework was evaluated on the Vindr-Mammo and CBIS-DDSM datasets using five independent data splits.

3.1 Classification Performance

Mammo-Clustering achieved state-of-the-art results, with an AUC of 0.828 on Vindr-Mammo (a 3.5% improvement over the next best model) and 0.805 on CBIS-DDSM (a 2.5% improvement). These results demonstrate the superior feature extraction capabilities of the clustering-based approach.

3.2 Weakly Supervised Localization

A significant advantage of this framework is its ability to perform accurate lesion localization using only image-level labels. The model achieved a high recall of 68.5% and the lowest Miss Detection Rate (29.4%) among compared models. Visualizations confirm that the CoC-based activation maps are sharper and more precisely aligned with actual lesions compared to CNN or ViT alternatives.

Qualitative visualization of lesion localization and activation maps across different lesion types (Mass, Calcification, and Both).

4. Discussion and Conclusion

Mammo-Clustering demonstrates that context clustering is a powerful paradigm for high-resolution medical imaging. By treating images as point sets, the model avoids the pitfalls of traditional down-sampling and better captures the biological characteristics of breast lesions. While current limitations include a fixed number of patches and challenges with class imbalance, future work will explore dynamic routing for patch selection. Overall, this framework offers a scalable, accurate solution for large-scale breast cancer screening.

Summary updated

Abstract

Breast cancer is a significant global health issue, and the diagnosis of breast cancer through imaging remains challenging. Mammography images are characterized by extremely high resolution, while lesions often occupy only a small portion of the image. Down-sampling in neural networks can easily lead to the loss of microcalcifications or subtle structures. To tackle these challenges, we propose a Context Clustering based triple information fusion framework. First, in comparison to CNNs or transformers, we observe that Context clustering methods are (1) more computationally efficient and (2) better at associating structural or pathological features. This makes them particularly well-suited for mammography in clinical settings. Next, we propose a triple information fusion mechanism that integrates global, feature-based local, and patch-based local information. The proposed approach is rigorously evaluated on two public datasets, Vindr-Mammo and CBIS-DDSM, using five independent data splits to ensure statistical robustness. Our method achieves an AUC of 0.828 \pm 0.020 on Vindr-Mammo and 0.805 \pm 0.020 on CBIS-DDSM, outperforming the second best method by 3.5% and 2.5%, respectively. These improvements are statistically significant (p&lt;0.05), highlighting the advantages of the Context Clustering Network with triple information fusion. Overall, our Context Clustering framework demonstrates strong potential as a scalable and cost-effective solution for large-scale mammography screening, enabling more efficient and accurate breast cancer detection. Access to our method is available at https://github.com/Sohyu1/Mammo-Clustering.
BibTeX
@misc{yang2025mammo,
  author = {Shilong Yang and Chulong Zhang and Xiaokun Liang and Qi Zang and Juan Yu and Liang Zeng and Xiao Luo and Yadi Xing and Tingting Peng and Qi Li and Linlin Shen and Yaoqin Xie},
  title = {{Mammo-Clustering: Context Clustering based Multi-view Tri Level Information Fusion for Lesion Location and Classification in Mammography}},
  year = {2025},
  eprint = {2507.18642},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.18642},
  howpublished = {arXiv},
}
Figure