A Multi-scale Linear-time Encoder for Whole-Slide Image Analysis

We introduce Multi-scale Adaptive Recurrent Biomedical Linear-time Encoder (MARBLE), the first \textit{purely Mamba-based} multi-state multiple instance learning (MIL) framework for whole-slide image (WSI) analysis. MARBLE processes multiple magnification levels in parallel and integrates coarse-to-fine reasoning within a linear-time state-space model, efficiently capturing cross-scale dependencies with minimal parameter overhead. WSI analysis remains challenging due to gigapixel resolutions and hierarchical magnifications, while existing MIL methods typically operate at a single scale and transformer-based approaches suffer from quadratic attention costs. By coupling parallel multi-scale processing with linear-time sequence modeling, MARBLE provides a scalable and modular alternative to attention-based architectures. Experiments on five public datasets show improvements of up to \textbf{6.9%} in AUC, \textbf{20.3%} in accuracy, and \textbf{2.3%} in C-index, establishing MARBLE as an efficient and generalizable framework for multi-scale WSI analysis.

💡 Research Summary

The paper introduces MARBLE (Multi‑scale Adaptive Recurrent Biomedical Linear‑time Encoder), a novel multiple‑instance learning (MIL) framework designed specifically for whole‑slide image (WSI) analysis in computational pathology. Traditional MIL approaches either operate on a single magnification level or rely on transformer architectures whose quadratic attention cost makes scaling to gigapixel images prohibitive. MARBLE addresses both limitations by (1) employing a Mamba‑based structured state‑space model (SSM) that processes token sequences in linear time, and (2) processing multiple magnification levels in parallel while fusing coarse‑to‑fine information through a lightweight, token‑aligned conditioning mechanism.

Each magnification level k (k = 0 … S, where 0 is the coarsest) is tiled into non‑overlapping patches, and a frozen 1024‑dimensional embedding (extracted by the publicly released UNI model) represents each patch. For every level a single Mamba‑2 block (depth = 1, model dimension = 1024) encodes the token sequence with O(T_k · D) complexity, where T_k is the number of tokens at that level. The parallel execution means the wall‑clock time is dominated by the longest (finest) sequence, preserving linear scalability.

The key innovation is the coarse‑to‑fine fusion. Before encoding level k > 0, each fine‑scale token x_i^{(k)} is concatenated with its parent token y_{p_k(i)}^{(k‑1)} from the coarser level, where p_k(i) is a deterministic mapping derived from the spatial grid (each fine patch lies entirely within a single coarse patch). A linear projection ϕ^{(k)} maps the concatenated vector back to the model dimension, yielding an enriched token \tilde{x}_i^{(k)}. This operation adds only a single projection matrix per level, incurring negligible parameter overhead while allowing the fine‑scale representation to be conditioned on global context.

After all levels are encoded, the finest‑scale token set S is aggregated using attention pooling. A learnable query vector w produces attention weights α(y) = exp(wᵀy)/∑exp(wᵀy′), and the slide‑level embedding z = ∑_{y∈S} α(y) · y is obtained. For classification, a linear layer maps z to logits and cross‑entropy loss is minimized. For survival analysis, a Cox proportional‑hazards head computes a risk score r = βᵀz and optimizes the negative partial log‑likelihood with an ℓ₂ regularizer.

Two regularization strategies further improve robustness. Random coarse‑branch drop randomly removes a fraction α (set to 0.1 after a small hyper‑parameter sweep) of the coarsest tokens and all their descendants during training, encouraging the model to rely on multiple scales. Scan‑order neutrality randomly permutes token order within each level before encoding, eliminating implicit positional bias; because parent‑child lookups are explicit, the model remains permutation‑invariant.

The authors evaluate MARBLE on five public datasets: two classification tasks (PANDA prostate cancer and TCGA‑NSCLC lung cancer) and three survival tasks (TCGA KIRP, LUAD, and STAD). Slides are tiled into 256 × 256 patches, and both available magnifications (e.g., 10× and 40×) are processed simultaneously. Baselines include a broad spectrum of MIL methods—ABMIL, CLAM, DSMIL, TransMIL, S4‑MIL, DTFD‑MIL—as well as recent Mamba‑based models (MambaMIL, SRMambaMIL, 2DMambaMIL). All experiments use identical patch embeddings, training hyper‑parameters (AdamW, learning rate 3e‑5, cosine decay, 30 epochs, early stopping) and are repeated five times with fixed random seeds.

Results show consistent superiority of MARBLE. On PANDA, MARBLE achieves 71.0 % accuracy (+20.3 pp over the best baseline) and 0.8878 AUC (+6.94 pp). On TCGA‑NSCLC, it reaches 0.9730 AUC, surpassing all transformer and state‑space baselines. In survival analysis, MARBLE attains the highest C‑index across all three cohorts (0.8184 for KIRP, 0.6432 for LUAD, 0.6510 for STAD), improving up to 2.3 pp over the next best method. Ablation studies comparing single‑scale (coarse‑only or fine‑only) versus the combined two‑scale configuration demonstrate that the fusion of both scales yields the best performance on both classification and survival tasks, confirming that coarse global context stabilizes reasoning while fine detail provides discriminative power.

In summary, MARBLE delivers a scalable, parameter‑efficient solution for multi‑magnification WSI analysis by integrating (1) linear‑time SSM encoding, (2) lightweight cross‑scale conditioning, and (3) robust regularization. The framework eliminates the quadratic bottleneck of attention‑based MIL, captures rich hierarchical information, and sets new performance benchmarks on several pathology datasets. Future work will explore data‑driven token traversal, selective patch routing for further efficiency, and extending the architecture to three or more native magnifications.