@Article{cmes.2026.084724,
AUTHOR = {Muhammad Zaheer Sajid, Muhammad Fareed Hamid, Nauman Ali Khan, Imran Qureshi},
TITLE = {WaSA-Net: Wavelet-Guided Tokenization and Dynamic Sparse Attention for Histopathology Image Classification},
JOURNAL = {Computer Modeling in Engineering \& Sciences},
VOLUME = {148},
YEAR = {2026},
NUMBER = {1},
PAGES = {--},
URL = {http://www.techscience.com/CMES/v148n1/68220},
ISSN = {1526-1506},
ABSTRACT = {Digital pathology is rapidly transforming histopathological diagnosis, yet many existing deep learning models treat all spatial regions uniformly and do not exploit the multi-frequency structure of tissue, which limits both diagnostic accuracy and computational efficiency. This paper proposes WaSA-Net, an end-to-end architecture that integrates three complementary modules for histopathological image analysis. First, the Wavelet-Guided Tokenization (WGT) module decomposes input images into frequency-aware representations using learnable wavelet-like filters, so that both global tissue structures and fine-grained cellular patterns are exposed to attention from the first layer. Second, the Dynamic Sparse Attention with Pathology Priors (DSA-PP) module adaptively selects diagnostically informative tokens through a lightweight gating mechanism and incorporates learnable pathology prior tokens that embed domain-specific inductive biases, reducing attention complexity while preserving critical contextual information. Third, the Cross-Frequency Feature Pyramid Fusion (CFFPF) module performs bidirectional cross-attention across frequency bands and applies adaptive per-sample frequency weighting to identify the most discriminative frequency components for each tissue type. The proposed architecture is evaluated on three widely used histopathology benchmarks: PatchCamelyon for metastasis detection, PathMNIST for multi-class colorectal tissue classification, and BreakHis for breast cancer diagnosis. WaSA-Net achieves strong performance with only 4.8M parameters, reaching 95.91% accuracy (AUC 0.9981) on PathMNIST, 93.47% accuracy (AUC 0.9812) on PatchCamelyon, and 96.72% accuracy (AUC 0.9923) on BreakHis. Despite its compact design, WaSA-Net matches or surpasses larger models while requiring no external pre-training data. These results indicate that frequency-aware representations and dynamic sparse attention can improve both efficiency and diagnostic performance in digital pathology.},
DOI = {10.32604/cmes.2026.084724}
}