Research map / Image generation / Fast diffusion sampling
L-Diffusion: Laplace Diffusion for Efficient Pathology Image Segmentation
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 L-Diffusion, which models pathology image components with Laplace noise in a diffusion process and refines per-pixel latent trajectory vectors with contrastive learning before a ConvNeXt segmentation head.
Problem
Pathology image segmentation suffers from costly gigapixel annotation and poor accuracy on tail (rare) tissue and cell classes caused by long-tail class distributions.
Approach
Replace the Gaussian noise in diffusion with multiple Laplace distributions, one per component, so that component distributions separate more across diffusion steps; then apply pixel latent vector contrastive learning over the per-pixel sequence of diffusion feature maps and segment with ConvNeXt.
Key results
- On tissue segmentation, L-Diffusion reports DICE of 82.38 on CRCD, 86.47 on PUMA and 91.77 on BCSS (per Table 1), with mIoU improvements of up to 16.52 points.
- On cell segmentation (Table 2), L-Diffusion reaches DICE 96.11 on CRCD, a gain of +20.09 over the best baseline as stated in the paper.
- Replacing Gaussian with Laplace noise plus contrastive learning raises average tissue DICE from 17.41 (Gaussian with contrastive) to 85.75 (Laplace with contrastive) in ablation (Table 4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Laplace distributions instead of Gaussian distributions (ablation) | L-Diffusion · Three tissue datasets (average) | DICE: 17.41 -> 85.75 | Helped |
| Add pixel latent vector contrastive learning (ablation) | L-Diffusion with Laplace distributions · Three tissue datasets (average) | DICE: 26.02 -> 85.75 | Helped |
| Use Laplace distributions and contrastive learning together (ablation) | L-Diffusion · Three tissue datasets (average) | FwIoU: 23.26 -> 89.74 | Helped |
| Increase the annotation ratio from 10% to 30% (ablation) | L-Diffusion · PUMA tissue | DICE: 19.95 -> 83.51 | Helped |
| Increase the annotation ratio from 30% to 100% (ablation) | L-Diffusion · PUMA tissue | DICE: 83.51 -> 92.11 | Helped |
| Replace Laplace noise with Cauchy noise (ablation) | L-Diffusion · Three tissue datasets (average) | DICE: 85.75 -> 16.25 | Hurt |
| Replace Laplace noise with Student’s t noise (ablation) | L-Diffusion · Three tissue datasets (average) | Runtime: 8882 -> 20440 | Hurt |
| Use L-Diffusion rather than the strongest baseline on each tissue benchmark | Pathology tissue segmentation methods · PUMA tissue | DICE: +7.16 | Helped |
| Use L-Diffusion rather than the strongest baseline on each tissue benchmark | Pathology tissue segmentation methods · CRCD tissue | mIoU: +16.52 | Helped |
| Use L-Diffusion rather than the strongest baseline on each cellular benchmark | Pathology cellular segmentation methods · CRCD cellular | DICE: +20.09 | Helped |
| Use L-Diffusion rather than the strongest baseline on each cellular benchmark | Pathology cellular segmentation methods · PanNuke cellular | mIoU: +14.42 | Helped |
| Use L-Diffusion rather than the strongest baseline | Remote-sensing image segmentation methods · Massachusetts-Building | IoU: 76.41 -> 80.30 | Helped |
Surprising findings
- Laplace noise without contrastive learning yields very low tissue DICE (26.02), showing contrastive learning is essential alongside the distribution change.
- Cauchy noise gives a much lower DICE (16.25) than Gaussian-style alternatives, attributed to undefined expectation and variance.
Limitations
- Contrastive learning loss requires annotated images, so it applies only to annotated subset.
- Segmentation module needs a sufficient number of annotated samples, so annotation demand remains.
- Hardware and training compute are not reported.
Experimental setup
Datasets: CRCD PUMA BCSS PanNuke Massachusetts-Building
Benchmarks: CRCD PUMA BCSS PanNuke Massachusetts-Building
Metrics: DICE MPA mIoU FwIoU IoU F1 Precision
Baselines: FastFCN U-Net++ Swin-UNet SAMUS SAMed SAMPath UN-SAM DeepLabv3 DeepLabv3+ DenseASPP BONUS
Builds on: DDPM ConvNeXt SAMPath BONUS
Closest papers by meaning
- MedSegDiff-V2: Diffusion-Based Medical Image Segmentation with Transformer (AAAI 2024 · similarity 0.86)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.86)
- Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction (NeurIPS 2026 · similarity 0.86)
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.85)
- EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models (ICLR 2024 · similarity 0.85)
- LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models (ICLR 2025 · similarity 0.85)
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.