MambaMIL+: Modeling Long-Term Contextual Patterns for Gigapixel Whole Slide Image

Whole-slide images (WSIs) are an important data modality in computational pathology, yet their gigapixel resolution and lack of fine-grained annotations challenge conventional deep learning models. Multiple instance learning (MIL) offers a solution by treating each WSI as a bag of patch-level instances, but effectively modeling ultra-long sequences with rich spatial context remains difficult. Recently, Mamba has emerged as a promising alternative for long sequence learning, scaling linearly to thousands of tokens. However, despite its efficiency, it still suffers from limited spatial context modeling and memory decay, constraining its effectiveness to WSI analysis. To address these limitations, we propose MambaMIL+, a new MIL framework that explicitly integrates spatial context while maintaining long-range dependency modeling without memory forgetting. Specifically, MambaMIL+ introduces 1) overlapping scanning, which restructures the patch sequence to embed spatial continuity and instance correlations; 2) a selective stripe position encoder (S2PE) that encodes positional information while mitigating the biases of fixed scanning orders; and 3) a contextual token selection (CTS) mechanism, which leverages supervisory knowledge to dynamically enlarge the contextual memory for stable long-range modeling. Extensive experiments on 20 benchmarks across diagnostic classification, molecular prediction, and survival analysis demonstrate that MambaMIL+ consistently achieves state-of-the-art performance under three feature extractors (ResNet-50, PLIP, and CONCH), highlighting its effectiveness and robustness for large-scale computational pathology

💡 Research Summary

**
The paper introduces MambaMIL+, a novel multiple‑instance learning (MIL) framework designed for gigapixel whole‑slide images (WSIs) in computational pathology. While recent state‑space models such as Mamba have demonstrated linear‑time scalability for extremely long token sequences, two critical shortcomings limit their effectiveness for WSIs: (1) a lack of explicit spatial‑context modeling, and (2) exponential memory decay that causes early tokens to lose influence as the sequence length grows to tens of thousands of patches.

To overcome these issues, the authors propose three complementary components. First, overlapping scanning extracts patches with a predefined overlap (e.g., 30‑50 %). Overlapping patches share pixel information, thereby embedding spatial continuity directly into the token sequence and mitigating the discontinuity introduced by the conventional non‑overlapping, i.i.d. assumption. Second, a Selective Stripe Position Encoder (S2PE) replaces the naïve row‑major positional encoding used in prior 2‑D Mamba variants. S2PE encodes positions stripe‑wise (horizontal and vertical strips) and selectively activates positional tokens, reducing the bias of a fixed scanning order while preserving global spatial relationships. Third, the Contextual Token Selection (CTS) mechanism leverages slide‑level supervision to dynamically identify and retain high‑value tokens in a dedicated long‑term memory buffer. By preserving the hidden states of these salient tokens, CTS counteracts the exponential decay inherent in the discretized transition matrix of Mamba, ensuring that early, diagnostically important regions continue to contribute throughout the forward pass.

The overall architecture integrates these modules into a State‑Space Duality (SSD)‑based Mamba backbone. SSD simplifies the transition matrix to a scalar, achieving linear computational complexity while still supporting selective information propagation. The pipeline proceeds as follows: (1) patches are extracted with overlap and encoded by a pretrained feature extractor (ResNet‑50, PLIP, or CONCH); (2) the resulting embeddings are reordered according to the overlapping scan; (3) S2PE injects spatial positional information; (4) the token stream passes through the SSD‑Mamba layers; (5) CTS dynamically selects tokens for long‑term storage; (6) a bag‑level aggregator (e.g., attention‑based pooling) produces the final slide prediction.

Extensive experiments were conducted on 20 public benchmarks covering diagnostic classification, molecular alteration prediction, and survival analysis. Across all tasks and three feature extractors, MambaMIL+ consistently outperformed 11 state‑of‑the‑art MIL methods, including TransMIL, CLAM, and S4MIL. Average improvements ranged from 2 % to 5 % absolute AUROC/AUPRC gains, with the most pronounced benefits observed on tasks requiring fine‑grained spatial reasoning. Ablation studies confirmed that each component contributes uniquely: overlapping scanning alone yields a 3 % AUROC boost, S2PE further refines spatial awareness, and CTS prevents performance collapse on sequences longer than ~30 k tokens, especially in survival prediction where the concordance index drops sharply without CTS.

In summary, MambaMIL+ demonstrates that (i) spatial context can be effectively re‑introduced into token‑based MIL pipelines via overlapping scans, (ii) bias‑free positional encoding can be achieved with stripe‑wise selective encoding, and (iii) supervised token selection can mitigate memory decay in linear‑time state‑space models. The work establishes Mamba as a powerful backbone for computational pathology and opens avenues for future research on more sophisticated token‑selection strategies and multimodal integration of histopathology with clinical data.