Research map / Image generation / Fast diffusion sampling
MExD: An Expert-Infused Diffusion Model for Whole-Slide Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MExD, a WSI classification model that pairs a sparse mixture-of-experts patch aggregator (Dyn-MoE) with a diffusion classifier (Diff-C) that generates slide-level class predictions conditioned on expert insights and a prior prediction.
Problem
Whole-slide image classification in multiple instance learning suffers from noisy non-informative patches, class imbalance between positive and negative instances, and weak feature aggregation under discriminative paradigms.
Approach
Replaces discriminative MIL classification with a conditional generative diffusion process: a MoE aggregator with one negative and K positive experts sparsifies patches and yields a prior prediction plus expert confidences, which condition a DDPM-style denoiser (Diff-C) that outputs the bag label.
Key results
- On Camelyon16 with CTransPath features, MExD reaches 97.29% F1, 97.48% ACC and 98.87% AUC, versus 96.10% F1 for PAMIL.
- On BRACS with ViT (MoCo v3) features, MExD reaches 80.10% F1 and 81.01% ACC, improving over IBMIL's 74.74% F1.
- Dyn-MoE adds about 2.79% average gain and Diff-C a further 1.45% on TCGA-NSCLC and BRACS in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Dyn-MoE to the Adapter + MLP-C baseline (ablation) | MExD · TCGA-NSCLC | F1-score: 91.37 -> 94.28; +2.89 | Helped |
| replace MLP-C with Diff-C without Dyn-MoE (ablation) | MExD baseline · TCGA-NSCLC | F1-score: 91.37 -> 92.37; +1.00 | Helped |
| add Diff-C after adding Dyn-MoE (ablation) | MExD · TCGA-NSCLC | F1-score: 94.28 -> 96.66; +2.38 | Helped |
| use Dyn-MoE and Diff-C in place of the baseline configuration (ablation) | MExD · BRACS | F1-score: 68.59 -> 75.95 | Helped |
| set α0=0.25 and α1=0.50 (ablation) | MExD · BRACS | F1-score: 77.77 -> 80.33 | Helped |
| reduce α1 from 0.50 to 0.10 with α0=0.25 (ablation) | MExD · BRACS | F1-score: 80.33 -> 74.48 | Hurt |
| remove prior prediction ρθ (ablation) | MExD · TCGA-NSCLC | F1-score: 93.80; +0.47% reported on TCGA-DSCLC | Hurt |
| remove prior prediction ρθ (ablation) | MExD · BRACS | F1-score: 75.56; +1.18% | Hurt |
| increase denoising steps from 100 to 200 (ablation) | MExD with prior prediction · BRACS | F1-score: 77.74 -> 80.15 | Helped |
| increase denoising steps from 200 to 300 (ablation) | MExD with prior prediction · BRACS | F1-score: 80.15 -> 80.33 | No clear effect |
| increase denoising steps from 300 to 400 (ablation) | MExD with prior prediction · BRACS | F1-score: 80.33 -> 80.15 | No clear effect |
| increase denoising steps from 100 to 200 (ablation) | MExD without prior prediction · BRACS | F1-score: 75.56 -> 77.74 | Helped |
Surprising findings
- Removing the prior prediction rho_theta slightly lowers performance but requires more denoising steps (400 vs 200) to reach similar results.
Limitations
- Evaluated only on three WSI benchmarks with binary and multi-class subtyping tasks.
- Relies on frozen offline patch features from pretrained extractors, so the feature extractor is not trained end to end.
Experimental setup
Datasets: Camelyon16 TCGA-NSCLC BRACS
Benchmarks: Camelyon16 TCGA-NSCLC BRACS
Metrics: F1-score accuracy AUC PAvPU
Baselines: ABMIL DSMIL CLAM-SB CLAM-MB TransMIL DTFD-MIL MHIM-MIL MambaMIL ACMIL IBMIL WiKG PAMIL
Builds on: CARD TransMIL denoising diffusion probabilistic model MoCo v3 IBMIL
Closest papers by meaning
- FAST: A Dual-tier Few-Shot Learning Paradigm for Whole Slide Image Classification (NeurIPS 2024 · similarity 0.87)
- OODML: Whole Slide Image Classification Meets Online Pseudo-Supervision and Dynamic Mutual Learning (AAAI 2025 · similarity 0.85)
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.85)
- Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction (CVPR 2024 · similarity 0.85)
- Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification (AAAI 2025 · similarity 0.84)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.