Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DeltaMIL: Gated Memory Integration for Efficient and Discriminative Whole Slide Image Analysis

Yueting Zhu, Yuehao Song, Shuai Zhang, Wenyu Liu, Xinggang Wang

TL;DR

DeltaMIL tackles the challenge of discriminating informative regions in ultra-high-resolution WSIs by introducing a gated delta memory mechanism that selectively forgets outdated information and writes new, correlated signals. A locality-aware delta rule fuses local 2D spatial context with global patch representations, enabling robust, discriminative slide-level analysis. The method achieves state-of-the-art performance on both survival prediction and slide-level classification across diverse datasets and feature extractors, while ablations confirm the complementary value of its local, gated, and delta components. The work offers strong generalizability and practical potential for computational pathology by improving robustness to redundant information and enhancing interpretability through region-focused attention.

Abstract

Whole Slide Images (WSIs) are typically analyzed using multiple instance learning (MIL) methods. However, the scale and heterogeneity of WSIs generate highly redundant and dispersed information, making it difficult to identify and integrate discriminative signals. Existing MIL methods either fail to discard uninformative cues effectively or have limited ability to consolidate relevant features from multiple patches, which restricts their performance on large and heterogeneous WSIs. To address this issue, we propose DeltaMIL, a novel MIL framework that explicitly selects semantically relevant regions and integrates the discriminative information from WSIs. Our method leverages the gated delta rule to efficiently filter and integrate information through a block combining forgetting and memory mechanisms. The delta mechanism dynamically updates the memory by removing old values and inserting new ones according to their correlation with the current patch. The gating mechanism further enables rapid forgetting of irrelevant signals. Additionally, DeltaMIL integrates a complementary local pattern mixing mechanism to retain fine-grained pathological locality. Our design enhances the extraction of meaningful cues and suppresses redundant or noisy information, which improves the model's robustness and discriminative power. Experiments demonstrate that DeltaMIL achieves state-of-the-art performance. Specifically, for survival prediction, DeltaMIL improves performance by 3.69\% using ResNet-50 features and 2.36\% using UNI features. For slide-level classification, it increases accuracy by 3.09\% with ResNet-50 features and 3.75\% with UNI features. These results demonstrate the strong and consistent performance of DeltaMIL across diverse WSI tasks.

DeltaMIL: Gated Memory Integration for Efficient and Discriminative Whole Slide Image Analysis

TL;DR

DeltaMIL tackles the challenge of discriminating informative regions in ultra-high-resolution WSIs by introducing a gated delta memory mechanism that selectively forgets outdated information and writes new, correlated signals. A locality-aware delta rule fuses local 2D spatial context with global patch representations, enabling robust, discriminative slide-level analysis. The method achieves state-of-the-art performance on both survival prediction and slide-level classification across diverse datasets and feature extractors, while ablations confirm the complementary value of its local, gated, and delta components. The work offers strong generalizability and practical potential for computational pathology by improving robustness to redundant information and enhancing interpretability through region-focused attention.

Abstract

Whole Slide Images (WSIs) are typically analyzed using multiple instance learning (MIL) methods. However, the scale and heterogeneity of WSIs generate highly redundant and dispersed information, making it difficult to identify and integrate discriminative signals. Existing MIL methods either fail to discard uninformative cues effectively or have limited ability to consolidate relevant features from multiple patches, which restricts their performance on large and heterogeneous WSIs. To address this issue, we propose DeltaMIL, a novel MIL framework that explicitly selects semantically relevant regions and integrates the discriminative information from WSIs. Our method leverages the gated delta rule to efficiently filter and integrate information through a block combining forgetting and memory mechanisms. The delta mechanism dynamically updates the memory by removing old values and inserting new ones according to their correlation with the current patch. The gating mechanism further enables rapid forgetting of irrelevant signals. Additionally, DeltaMIL integrates a complementary local pattern mixing mechanism to retain fine-grained pathological locality. Our design enhances the extraction of meaningful cues and suppresses redundant or noisy information, which improves the model's robustness and discriminative power. Experiments demonstrate that DeltaMIL achieves state-of-the-art performance. Specifically, for survival prediction, DeltaMIL improves performance by 3.69\% using ResNet-50 features and 2.36\% using UNI features. For slide-level classification, it increases accuracy by 3.09\% with ResNet-50 features and 3.75\% with UNI features. These results demonstrate the strong and consistent performance of DeltaMIL across diverse WSI tasks.
Paper Structure (28 sections, 9 equations, 7 figures, 8 tables)

This paper contains 28 sections, 9 equations, 7 figures, 8 tables.

Table of Contents

  1. Introduction
  2. RELATED WORK
  3. MIL for Whole Slide Images
  4. Efficient Sequence Modeling
  5. METHOD
  6. Motivation
  7. Overall Architecture
  8. GDNBlock
  9. Locality-Aware Gated Delta Rule
  10. Architecture Details
  11. EXPERIMENTS
  12. Datasets
  13. Implementation Details
  14. Results
  15. Survival Prediction
  16. ...and 13 more sections

Figures (7)

  • Figure 1: Comprehensive comparison across diverse whole-slide image (WSI) analysis tasks, showing strong and well-balanced performance.
  • Figure 2:
  • Figure 3: Accuracy variation with respect to the retained patch ratio. This figure presents the changes in model ACC on the NSCLC dataset (panels (a)–(c)) and the BRACS bracs dataset (panels (d)–(f)) as the proportion of retained patches decreases. Each set of panels corresponds to the three sampling strategies, Random-k, Top-k, and Bottom-k, applied to ABMIL abmil, MambaMIL mambamil, and Mamba2MIL mamba2mil. The pronounced differences among the three strategies indicate that the model’s decision-making relies primarily on a tiny subset of highly informative patches, while most patches are irrelevant to the final prediction, demonstrating the inherent redundancy in pathology images.
  • Figure 4: Overall framework of the proposed model. The left part shows the overall workflow: WSIs are processed by a feature extractor and then pass through $L$ stacked Gated DeltaNet blocks, followed by an aggregator to produce slide-level predictions. The right part details the Gated DeltaNet module, which enhances the original Gated DeltaNet by integrating local features reconstructed from the 2D spatial layout. This design enables joint modeling of long-range dependencies and fine-grained spatial patterns in histopathology data.
  • Figure 5: Illustration of the Gated Delta rule, which consists of two operations: (a) Remove Old — discarding outdated information through a decay mechanism, and (b) Write New — updating with new information.
  • ...and 2 more figures