2026

DSNet: A lightweight CNN for cross-magnification whole slide image analysis for breast cancer diagnosis

Md Rofiqul Bari, Jannatul Ferdaus, Amir Sohel, Md Alamgir Kabir, Md Shakhawat Hossain

Knowledge-Based Systems , Vol. 351 , pp. 116753

View Publication DOI: 10.1016/j.knosys.2026.116753

DSNet: A lightweight CNN for cross-magnification whole slide image analysis for breast cancer diagnosis

Abstract

Convolutional neural networks (CNNs) support automated breast cancer WSI analysis but often require substantial computational resources. We propose DSNet, a lightweight and interpretable CNN with only 1.31 million parameters (5.01 MB). DSNet achieved accuracies of 99.8%, 94.9%, 92.7%, and 88.0% on BreakHis at 400×, 200×, 100×, and 40× magnifications, respectively, and 90.5% on the external BACH dataset. It also achieved the fastest inference time (51.8 ms) and highest accuracy-to-GFLOPs ratio (81.3) among the evaluated models. Score-CAM highlighted diagnostically relevant regions, supporting model interpretability. Overall, DSNet offers an efficient, robust, and interpretable framework for breast cancer WSI analysis.

Citation

Md Rofiqul Bari, Jannatul Ferdaus, Amir Sohel, Md Alamgir Kabir, Md Shakhawat Hossain. "DSNet: A lightweight CNN for cross-magnification whole slide image analysis for breast cancer diagnosis." Knowledge-Based Systems 351 (2026): 116753.

BibTeX

@article{pub54_2026,
  title={DSNet: A lightweight CNN for cross-magnification whole slide image analysis for breast cancer diagnosis},
  author={Md Rofiqul Bari, Jannatul Ferdaus, Amir Sohel, Md Alamgir Kabir, Md Shakhawat Hossain},
  journal={Knowledge-Based Systems},
  volume={351},
  pages={116753},
  year={2026},
  doi={10.1016/j.knosys.2026.116753}
}