MSCAnet: A multi-scale cross-attention network for chromosomal instability prediction in breast cancer whole-slide images
- 1Department of Intelligent Science and Technology, School of Automation, Northwestern Polytechnical University, Xi'an, Shaanxi 710129, China
- 2College of Artificial Intelligence, Henan Agricultural University, Zhengzhou, Henan 450046, China
- 3College of Artificial Intelligence, Henan Agricultural University, Zhengzhou, Henan 450046, China; Beijing Key Laboratory of Topological Statistics and Applications for Complex Systems, Beijing Institute of Mathematical Sciences and Applications, Beijing 101408, China; Corresponding author at: College of Artificial Intelligence, Henan Agricultural University, Zhengzhou, Henan 450046, China.
Abstract
Introduction: Chromosomal instability (CIN) is an important biomarker for breast cancer diagnosis, prognosis, and treatment stratification. However, conventional laboratory-based CIN assessment is often time-consuming and costly, while existing deep learning methods face challenges in integrating local cellular details with broader tissue context from whole-slide images (WSIs). Methods: We propose a multi-scale cross-attention neural network (MSCAnet) for CIN status prediction from hematoxylin and eosin (H&E)-stained breast cancer WSIs. MSCAnet employs dual patch-level branches to extract complementary nuclear and tissue structural features, while a CIN-aware cross-attention module dynamically integrates multi-scale information. The model was evaluated on The Cancer Genome Atlas Breast Invasive Carcinoma dataset and further assessed on an additional human epidermal growth factor receptor 2 (HER2) status prediction task. Results: MSCAnet achieved an accuracy of 77.46% and an area under the receiver operating characteristic curve of 0.809 for CIN prediction, demonstrating the highest numerical performance among the evaluated methods. However, the overlapping 95% confidence intervals with those of the strongest baselines precluded definitive conclusions regarding statistical superiority. Visualization analyses revealed spatial heterogeneity in CIN-associated tissue patterns. In the additional HER2 prediction task, MSCAnet achieved an accuracy of 82.72%, providing preliminary evidence of its applicability across molecular biomarker prediction tasks. Conclusions: MSCAnet demonstrates the potential of multi-scale cross-attention mechanisms for predicting molecular biomarkers from routine H&E-stained WSIs. Although the results are promising, further validation on larger, independent cohorts is required to establish the model's robustness, generalizability, and clinical utility.
Related
- Critical Fragility Emerges from Chromosomal Instability in Cancer
- p53 activity links chromosomal instability to pro-migratory extracellular vesicles
- Rapid Micronucleus-Sequencing (RaMN-Seq) Reveals Patterns of Bias in Induced and Spontaneous Chromosomal Instability
- Hypotension prediction index to predict hemodynamic instability related to renal replacement therapy in critically ill patients: A prospective observational study
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
BibTeX
@article{wang2026mscanet,
title = {{MSCAnet: A multi-scale cross-attention network for chromosomal instability prediction in breast cancer whole-slide images}},
author = {Ying Wang and Yue Hu and Liangliang Liu},
journal = {Intelligent Oncology},
year = {2026},
month = oct,
doi = {10.1016/j.intonc.2026.100091},
url = {https://doi.org/10.1016/j.intonc.2026.100091},
note = {CC BY},
}
RIS
TY - JOUR
TI - MSCAnet: A multi-scale cross-attention network for chromosomal instability prediction in breast cancer whole-slide images
AU - Ying Wang
AU - Yue Hu
AU - Liangliang Liu
PY - 2026
DA - 2026/10/25
JO - Intelligent Oncology
DO - 10.1016/j.intonc.2026.100091
UR - https://doi.org/10.1016/j.intonc.2026.100091
AB - Introduction: Chromosomal instability (CIN) is an important biomarker for breast cancer diagnosis, prognosis, and treatment stratification. However, conventional laboratory-based CIN assessment is often time-consuming and costly, while existing deep learning methods face challenges in integrating local cellular details with broader tissue context from whole-slide images (WSIs). Methods: We propose a multi-scale cross-attention neural network (MSCAnet) for CIN status prediction from hematoxylin and eosin (H&E)-stained breast cancer WSIs. MSCAnet employs dual patch-level branches to extract complementary nuclear and tissue structural features, while a CIN-aware cross-attention module dynamically integrates multi-scale information. The model was evaluated on The Cancer Genome Atlas Breast Invasive Carcinoma dataset and further assessed on an additional human epidermal growth factor receptor 2 (HER2) status prediction task. Results: MSCAnet achieved an accuracy of 77.46% and an area under the receiver operating characteristic curve of 0.809 for CIN prediction, demonstrating the highest numerical performance among the evaluated methods. However, the overlapping 95% confidence intervals with those of the strongest baselines precluded definitive conclusions regarding statistical superiority. Visualization analyses revealed spatial heterogeneity in CIN-associated tissue patterns. In the additional HER2 prediction task, MSCAnet achieved an accuracy of 82.72%, providing preliminary evidence of its applicability across molecular biomarker prediction tasks. Conclusions: MSCAnet demonstrates the potential of multi-scale cross-attention mechanisms for predicting molecular biomarkers from routine H&E-stained WSIs. Although the results are promising, further validation on larger, independent cohorts is required to establish the model's robustness, generalizability, and clinical utility.
SN - 2950-2616
SN - 2097-7131
ER -