Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Jagan Mohan Reddy Dwarampudi, Joshua Wong, Hien Van Nguyen, Tania Banerjee

TL;DR

This study tackles the challenge of scalable whole-slide image analysis by addressing cross-scale information with a linear-time, multi-scale MIL framework. MARBLE adopts parallel per-scale Mamba-2 encoders and a lightweight coarse-to-fine fusion that conditions fine-scale tokens on coarse-scale context, avoiding quadratic attention costs. The approach yields substantial improvements across five public datasets for both classification and survival analyses, with evidence that combining coarse and fine magnifications provides consistently better performance than either scale alone. Overall, MARBLE offers a scalable, generalizable alternative to transformer-based architectures for multi-scale WSI analysis with a compact parameter footprint and end-to-end trainability.

Abstract

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.

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

TL;DR

This study tackles the challenge of scalable whole-slide image analysis by addressing cross-scale information with a linear-time, multi-scale MIL framework. MARBLE adopts parallel per-scale Mamba-2 encoders and a lightweight coarse-to-fine fusion that conditions fine-scale tokens on coarse-scale context, avoiding quadratic attention costs. The approach yields substantial improvements across five public datasets for both classification and survival analyses, with evidence that combining coarse and fine magnifications provides consistently better performance than either scale alone. Overall, MARBLE offers a scalable, generalizable alternative to transformer-based architectures for multi-scale WSI analysis with a compact parameter footprint and end-to-end trainability.

Abstract

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.
Paper Structure (10 sections, 3 equations, 2 figures, 3 tables)

This paper contains 10 sections, 3 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Multi-Scale Encoding with Coarse-to-Fine Fusion
  4. Slide-Level Heads
  5. Regularization and Robustness
  6. Datasets and Baselines
  7. Results
  8. Conclusion
  9. Compliance with Ethical Standards
  10. Conflicts of Interest

Figures (2)

  • Figure 1: Overview of MARBLE. Left: a multi-resolution slide with three levels. Per-level sequences are formed and interleaved. Right: three Mamba-2 blocks $M_k$ process levels $k=0<1<2$. For $k>0$, each token is fused with its parent from level $k-1$. Final representations from $M_2$ are pooled into a slide-level embedding.
  • Figure 2: Effect of the drop regularizer rate $\alpha$ on test AUC for TCGA-NSCLC (10$\times$). $\alpha=0.1$ yields the best trade-off between regularization and retained discriminative signal.