Authors - Md. Riaz Mahmud, Kazi Asif Ahmed, Md. Rafiqul Islam, Kabya Guha Abstract - Modeling multi-scale spatial dependencies is essential in histopathology image analysis, where diagnostically relevant patterns span cellular textures and tissue-level structures. While convolutional neural networks effectively capture local features, they struggle to model long-range interactions, and transformer-based approaches address this limitation at the cost of quadratic computational complexity with respect to spatial resolution. In this work, we propose HiSS-Fuse, a linear-time hierarchical state-space fusion framework that integrates multi-scale fea ture representations using Mamba-based selective state-space modules. The proposed architecture progressively fuses local and global contex tual information across network depths while maintaining O(L) com putational complexity, where L denotes the number of spatial tokens. Experimental evaluation on the PathMNIST benchmark demonstrates that HiSS-Fuse achieves 97.0% classification accuracy with an AUC of 0.997 while maintaining strong computational efficiency. Ablation stud ies further confirm that hierarchical fusion systematically enhances rep resentation learning. Overall, HiSS-Fuse provides a scalable and compu tationally efficient alternative to quadratic attention-based architectures for multi-scale histopathology image analysis.
Open Zoom