.png){kind=icon}
.png){kind=icon}
Return
SSR-Mam2MIL: Spatial sequence reordering Mamba2 based multiple instance learning for computational pathology
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2025.109063.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multiple Instance Learning (MIL) is a promising paradigm for computational pathology based on Whole Slide Images (WSIs). Recently, building on the principles of State Space Models (SSMs), the Mamba model has been applied to MIL owing to its ability to model long sequences with linear complexity. Existing Mamba-based frameworks enhance dependency modeling through reordering operations. However, such operations disrupt spatial relationships among instances, resulting in reduced interpretability and suboptimal utilization of sequence information. To address this issue, a novel framework termed SSR-Mam2MIL is proposed to effectively model instance dependencies in WSIs. First, an Index-based Sequence Reordering Block (ISRB) is introduced, where instance sequences are reordered according to patch coordinates and cluster assignments to enable flexible scanning strategies. Second, an Adaptive Bidirectional Mamba2 Module (ABMM) is designed to capture long-range forward and backward dependencies and selectively fuse discriminative features. In addition, attention heatmaps are generated by leveraging the spatial interpretability of reordered instance sequences to validate the effectiveness of the proposed model. Extensive experiments demonstrated that the proposed model outperforms state-of-the-art MIL methods. On the Camelyon16 dataset, SSR-Mam2MIL achieves a binary classification area under the curve (AUC) of 78.15% and an accuracy (ACC) of 79.22%. On the NSCLC dataset, it achieves a tumor subtype classification AUC of 94.10% and an ACC of 87.83%. On the RCC dataset, it achieves an AUC of 98.68% and an ACC of 92.47%.
Journal
IF:
4.9
Papers:
9.8K
Citations:
2.4W
Organization
No organization information available