Research map / Image generation / Fast diffusion sampling
LeFusion: Controllable Pathology Synthesis via Lesion-Focused Diffusion Models
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 LeFusion, a lesion-focused diffusion model that generates lesion image-mask pairs by inpainting lesions into normal scans, with histogram-based texture control, multi-channel decomposition, and lesion mask diffusion (DiffMask), boosting nnUNet and SwinUNETR segmentation on lung nodule CT and cardiac MRI.
Problem
Real clinical pathology data is scarce and long-tailed; existing lesion synthesis methods struggle to keep backgrounds high-quality and offer limited control over lesion texture, mask, and class, which hurts downstream segmentation.
Approach
Train a diffusion model whose loss is restricted to the lesion mask, and at inference combine forward-diffused real background with reverse-diffused lesion foreground so only lesions are generated; add histogram conditioning for multi-peak textures, channel-wise decomposition for multi-class lesions, and a separate diffusion model for lesion masks controlled by bounding spheres.
Key results
- On LIDC with nnUNet, LeFusion-H with P+N' setting improved Dice from 78.26 (baseline) to 81.11 on synthetic data from copied masks; the full LeFusion-H+DiffMask pipeline reached 83.44 Dice.
- On Emidec, LeFusion-J+DiffMask (P+P'+N'') raised PMO Dice from 36.32 to 43.54 with nnUNet.
- Reported Dice gains of 5.18% (nnUNet) and 4.75% (SwinUNETR) on lung nodule segmentation over baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add histogram-based texture control to LeFusion (ablation) | LeFusion for lung nodule synthesis · LIDC, P+P', nnU-Net Dice | Dice: 78.77 -> 80.62 | Helped |
| Add histogram-based texture control to LeFusion (ablation) | LeFusion for lung nodule synthesis · LIDC, P+N', nnU-Net Dice | Dice: 77.67 -> 80.19 | Helped |
| Add multi-channel joint modeling of lesion classes (ablation) | LeFusion cardiac lesion synthesis · Emidec, P+P', nnU-Net PMO Dice | Dice: 34.79 -> 38.01 | Helped |
| Add multi-channel joint modeling of lesion classes (ablation) | LeFusion cardiac lesion synthesis · Emidec, P+P', SwinUNETR PMO Dice | Dice: 35.63 -> 37.99 | Helped |
| Add DiffMask-generated lesion masks (ablation) | LeFusion-H lung nodule synthesis · LIDC, P+N', nnU-Net Dice | Dice: 81.11 -> 82.66 | Helped |
| Add DiffMask-generated lesion masks (ablation) | LeFusion-J cardiac lesion synthesis · Emidec, P+N', nnU-Net PMO Dice | Dice: 37.31 -> 40.62 | Helped |
| Increase synthetic normal-scan data from N' to N'' (ablation) | LeFusion-H+DiffMask · LIDC, nnU-Net Dice | Dice: 82.66 -> 83.19 | Helped |
| Increase synthetic normal-scan data from N' to N'' (ablation) | LeFusion-J+DiffMask · Emidec, nnU-Net PMO Dice | Dice: 40.62 -> 42.44 | Helped |
| Add histogram-based texture control to LeFusion-J+DiffMask (ablation) | LeFusion-J+DiffMask · Emidec, P+P'+N'', nnU-Net MI Dice | Dice: 70.34 -> 71.28 | Helped |
| Use LeFusion-generated synthetic data instead of baseline-only training | nnU-Net · LIDC, P+P'+N'', Dice | Dice: 78.26 -> 83.44 | Helped |
| Use LeFusion-J+DiffMask synthetic data instead of baseline-only training | nnU-Net · Emidec, P+P'+N'', MI Dice | Dice: 68.61 -> 70.34 | Helped |
| Use LeFusion-J+DiffMask synthetic data instead of baseline-only training | nnU-Net · Emidec, P+P'+N'', PMO Dice | Dice: 36.32 -> 43.54 | Helped |
Surprising findings
- Without histogram control, generated lung nodules tend to match the lighter normal background, yielding overly subtle lesions.
- Several baseline synthesis methods (Hand-Crafted, Cond-Diffusion, RePaint) reduced downstream segmentation performance in several settings.
Limitations
- Histogram control helps less for cardiac lesions with small contrast differences, where LeFusion-J-H gives similar effects.
- Evaluated on only two datasets (lung nodule CT and cardiac MRI), both small.
Experimental setup
Datasets: LIDC-IDRI Emidec
Benchmarks: LIDC Emidec
Metrics: Dice NSD PSNR SSIM
Baselines: Copy-Paste Hand-Crafted lesion synthesis RePaint Cond-Diffusion Cond-Diffusion (L)
Builds on: RePaint Blended-Diffusion latent diffusion classifier-free cross-attention conditioning
Closest papers by meaning
- STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation (NeurIPS 2026 · similarity 0.85)
- L-Diffusion: Laplace Diffusion for Efficient Pathology Image Segmentation (ICML 2025 · similarity 0.85)
- Noise-Consistent Siamese-Diffusion for Medical Image Synthesis and Segmentation (CVPR 2025 · similarity 0.84)
- ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation (CVPR 2025 · similarity 0.84)
- Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.84)
- Beyond ViT Tokens: Masked-Diffusion Pretrained Convolutional Pathology Foundation Model for Cell-Level Dense Prediction (NeurIPS 2026 · 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.