PaperlayerView exact revision on arXiv

Source-linked AI summary

GWDM: A Guided Wavelet Diffusion Model for Realistic Histopathology Image Generation

Abdullah, Tao Huang, Ickjai Lee, Lei Bi, Euijoon Ahn

arXiv:2610.30775v1eess.IV

TL;DR

Limited, costly expert annotation and the challenge of preserving histomorphology and staining constrain histopathology image generation. GWDM combines wavelet diffusion with domain-specific measurements and ramped guidance losses. It reports consistent improvements in synthesis metrics across three benchmarks and positions the method as a basis for synthetic-data augmentation in computational pathology.

  • Problem

    Histopathology analysis is constrained by limited, costly expert-labeled data, and generic generation may fail to preserve structural and semantic consistency needed for tissue analysis.

    Source: §1 · pp. 1–2

  • Method

    GWDM conditions wavelet diffusion on morphological, spatial, and staining features, with ramped histomorphological guidance losses applied during training.

    Source: §1 · pp. 2–3

  • Results

    GWDM achieved consistent improvements in FID, KID, pFID, and path-FID across BCSS, MoNuSeg, and PAIP2019.

    Source: §8 · p. 12

  • Takeaways & Limitations

    GWDM offers a domain-aware approach to histopathology image generation and a foundation for synthetic-data augmentation in computational pathology workflows.

    Source: §7 · p. 12

    The current implementation operates at patch level and relies partly on handcrafted descriptors, which may limit scalability.

    Source: §7 · p. 12

Abstract

from arXiv · show

Histopathology images are recognized as a reliable and comprehensive modality for Deep Learning (DL) based diagnosis. However, effectiveness is constrained by limited large-scale annotated datasets, as collecting expert-labeled medical data is costly and time-intensive. Recent advances in generative models, particularly diffusion models, demonstrate strong capabilities in synthesizing medically relevant content for dataset augmentation. Existing generative frameworks treat histopathology patches as generic natural images, failing to enforce biological constraints such as histomorphological structures and heterogeneous staining patterns essential for valid synthesis. We propose a guided wavelet diffusion model (GWDM) framework that captures image details at multiple scales using wavelets and leverages controllable conditioning based on domain-specific histopathological features (e.g., tissue patterns and color variations). We further formulate a histomorphological guidance loss applied to real and generated patches. A loss ramping strategy gradually introduces these constraints during training, ensuring stable optimization and robust modeling of tissue architecture. GWDM was evaluated on three benchmark datasets (BCSS, MoNuSeg, and PAIP2019), consistently outperforming state-of-the-art generative methods. Ablation studies confirmed that domain-guided conditioning, perceptual loss, and structural similarity loss each contribute meaningfully to synthesis quality. Downstream nuclei and tissue segmentation experiments further validated the clinical utility of GWDM-generated images, demonstrating consistent performance improvements when synthetic data augments real training sets. This work shows that domain-specific conditioning and dual-loss optimization generate realistic histopathology patches, reducing data scarcity and improving clinical downstream segmentation performance.

1. Introduction

Histopathology analysis depends on large, expert-labeled datasets, but collecting them is costly, and realistic patch generation must preserve tissue structure and staining. GWDM combines wavelet-based generation with histomorphological guidance, and the paper reports improved synthesis quality and downstream segmentation performance.

1. Introduction

  • Large-scale annotated histopathology data are difficult and expensive to obtain, while tissue morphology and staining patterns make patch analysis challenging.The introduction describes the labor-intensive process of preparing and digitizing stained tissue slides.Source: §1 · p. 1
  • Wavelet diffusion can reduce dimensionality and inference time, but deeper decomposition risks losing critical information and domain-agnostic denoising can weaken structural and semantic consistency.The cited WDM approach reduces inference time by approximately fourfold compared with traditional DDPMs.Source: §1 · p. 2
  • The method conditions wavelet diffusion on morphological, multi-scale spatial, and staining measurements to model tissue architecture without paired data or spatial annotations.A ramping function gradually introduces the guidance constraints during training to support stable optimization and local-to-global structural consistency.Source: §1 · pp. 2–3
  • GWDM achieves higher synthesis quality than existing generative methods across three benchmark datasets, with ablations and downstream segmentation supporting its effectiveness and utility.The introduction names BCSS, MoNuSeg, and PAIP2019 but provides no metric values.Source: §1 · p. 3

2. RELATED WORK

Related work traces diffusion models from medical imaging to histopathology synthesis and domain-guided conditioning. These approaches motivate GWDM’s use of domain knowledge to steer generation.

2.1. Diffusion Model in Medical Imaging

  • Medical-imaging diffusion work spans reconstruction, cross-modal synthesis, translation, and synthetic-image augmentation, including reports of improved classifier accuracy when synthetic skin images are combined with real data.The cited examples include radiology, MRI, CT, and dermatology applications.Source: §2.1 · p. 3

2.2. Diffusion-Based Histopathology Image Generation

  • Histopathology diffusion methods condition synthesis on text, instance masks, unpaired masks, diagnostic text, morphology, color, or learned structural and textural features.Reported aims include tissue-structure control and improved downstream nuclei segmentation.Source: §2.2 · p. 3

2.3. Domain-Guided Generative Conditioning

  • Domain-guided diffusion uses signals such as spatial conditions, class prototypes, reports, anatomical masks, and genotype data to steer generation toward domain-specific targets.These approaches motivate integrating domain knowledge into diffusion models, as adopted in GWDM.Source: §2.3 · p. 3
  • GWDM’s framework decomposes patches into Haar wavelet coefficients, applies measurement-based histomorphology guidance during training, and generates patches through reverse diffusion and inverse wavelet transformation.The figure distinguishes training, which combines guidance with adversarial and reconstruction losses, from inference in the wavelet domain.Source: Figure 2 · p. 4

3. Method

GWDM formulates synthesis as matching domain-relevant histomorphological measurements between real and generated patches. It combines wavelet-domain diffusion with adversarial learning and training-time measurement guidance.

3.1. Problem Formulation

  • The goal is to synthesize patches whose staining, morphological, and spatial measurements match those of real patches, then assess utility through downstream segmentation.Source: §3.1 · p. 4

3.2. Proposed Method

  • GWDM unifies wavelet-domain diffusion with a GAN-based score learner and uses domain measurements to steer synthesis toward biologically faithful structure.Source: §3.2 · p. 4

3.2.1. Wavelet Representation

  • A discrete wavelet transform produces concatenated low- and high-frequency subbands, reducing spatial resolution fourfold for diffusion and adversarial training.Source: §3.2.1 · pp. 4–5

3.2.2. Training Stage

  • Training adds noise to real wavelet coefficients, then the generator maps a noisier state and latent noise to an estimated clean representation for adversarial denoising.The discriminator distinguishes real from generated denoising transitions conditioned on the noisier state.Source: §3.2.2 · p. 5

3.2.3. Measurement Model

  • The measurement operator combines morphological features, multi-scale spatial features, and H&E staining and optical-density characteristics.Source: §3.2.3 · p. 5

3.2.4. Histomorphological Guidance Losses

  • Guidance compares generated-image measurements with fixed reference measurements and combines this objective with adversarial denoising loss; epoch schedules ramp guidance up during early training.The text describes a stabilizer schedule that attenuates as training progresses.Source: §3.2.4 · p. 5

3.2.5. Inference Stage

  • At inference, the generator progressively denoises random Gaussian noise in the wavelet domain, samples previous states from the posterior, and maps the result back to RGB.No histomorphological guidance term is applied at inference.Source: §3.2.5 · pp. 5–6

4. EXPERIMENTS

The experiments evaluate GWDM on three histopathology datasets using fixed-size patches, compare its image fidelity with established metrics, and use common WDM-based implementation settings.

4.1. Datasets and Preprocessing

  • The evaluation covers the BCSS breast-cancer, MoNuSeg multi-organ nuclei, and PAIP2019 liver-cancer datasets, partitioned into 256×256-pixel patches.BCSS and PAIP2019 patches are non-overlapping; MoNuSeg images are partitioned into patches for training and testing.Source: §4.1 · p. 6, §4.1.1 · p. 6, §4.1.2 · p. 6

4.1.1. BCSS dataset

  • BCSS uses H&E-stained breast-cancer WSIs with ROIs annotated by 25 participants and validated by a senior pathologist.ROIs average 1.18 mm2 (SD: 0.80 mm2) at 40× magnification.Source: §4.1.1 · p. 6

4.1.2. MoNuSeg dataset

  • MoNuSeg contains 37 images with 21,623 expert-annotated nuclear boundaries spanning seven organ types and benign and malignant tissue.Images were acquired at 40× magnification from individual patients through TCGA.Source: §4.1.2 · p. 6

4.1.3. PAIP2019 dataset

  • PAIP2019 comprises H&E-stained liver-cancer WSIs scanned at 20× magnification and partitioned into non-overlapping 256×256-pixel patches.The source images range from 35,855 × 39,407 to 64,768 × 47,009 pixels.Source: §4.1.3 · p. 6

4.2. Implementation details

  • Training used a three-GPU cluster, while inference and evaluation used one GPU; all datasets used two sampling steps and 256×256-pixel images.Network settings followed the original WDM architecture.Source: §4.2 · p. 6

4.3. Evaluation metrics

  • Image-level evaluation used FID, KID, patch-FID, and path-FID, with lower scores indicating better agreement with the real-data distribution.patch-FID emphasizes local texture and structure; path-FID uses UNI features for a pathology-specific measure.Source: §4.3 · pp. 6–7

5. Results

GWDM generally outperforms the compared generative methods across the three datasets, while ablations report gains from guidance components and loss ramping. Qualitative comparisons and downstream segmentation figures also show benefits from synthetic-data augmentation.

5.1. Performance Comparison

  • GWDM outperforms existing models across datasets, with lower FID, KID, and pFID and competitive path-FID; on BCSS, its FID is 12.64 versus 90.74 for SPAD.On PAIP2019, GWDM improves over WDM on FID, KID, and pFID, but WDM has the lower path-FID; MoNuSeg gains are comparatively modest.Source: §5.1 · pp. 7–9
  • Figure 3 compares seven generative models on BCSS and MoNuSeg, highlighting texture preservation and cell-boundary accuracy in GWDM samples.The caption directs readers to Fig. 1 for ground-truth visual-quality images.Source: Figure 3 · p. 8
  • On MoNuSeg, GWDM’s pFID of 51.589 is comparable to DDGAN’s 50.373, while SPAD has a lower path-FID than GWDM, 624.464 versus 857.123.The passage attributes the narrower differentiation to MoNuSeg’s low-diversity, structurally homogeneous patches.Source: §5.1 · p. 8

5.2. Ablation Study

  • Figures 4 and 5 compare segmentation models on MoNuSeg and BCSS using real-only, 10k-synthetic-plus-2k-real, and 50k-synthetic-plus-2k-real training settings.Both captions report downstream performance improvements as more synthetic data is added.Source: Figure 4 · p. 9, Figure 5 · p. 10
  • With stain, spatial, and morphology guidance plus loss ramping, the MoNuSeg ablation reaches FID 45.02, a 12.6% reduction from the WDM baseline.The baseline FID is 51.49; the no-ramping variant reaches 46.42, while the individual guidance variants also reduce FID.Source: §5.2 · pp. 9–10

6. Downstream Segmentation

The downstream experiments test whether GWDM-generated images improve nuclei and tissue segmentation on MoNuSeg and BCSS. Across architectures, synthetic augmentation generally improves results, with gains varying by dataset and model.

6. Downstream Segmentation

  • On MoNuSeg and BCSS, adding GWDM-generated synthetic images generally improves segmentation across DeepLabV3+, UNet++, and SegFormer.The experiments compare real-only training with 10k or 50k synthetic images added to 2k real images.Source: §6 · p. 10
  • With 50k synthetic images on MoNuSeg, DeepLabV3+ Dice rises from 0.931 to 0.936 and mIoU from 0.872 to 0.882; on BCSS, its Dice rises from 0.738 to 0.745 and mIoU from 0.599 to 0.606.On MoNuSeg, UNet++ reaches Dice 0.956 and mIoU 0.917 with 50k synthetic images, compared with 0.954 and 0.914 using real data only.Source: §6 · p. 10
  • Qualitative results show more accurate masks, improved boundary delineation, and fewer false positives with synthetic augmentation, particularly in complex tissue regions.Figures include input images, ground-truth masks, predictions, and per-image Dice and IoU scores.Source: §6 · p. 11

7. Discussion

GWDM performs strongly across three histopathology benchmarks, and its synthetic patches improve downstream segmentation, particularly on difficult BCSS samples. The current patch-level implementation and partial reliance on handcrafted descriptors may limit scalability.

7. Discussion

  • GWDM’s measurement operator conditions denoising on morphology, multiscale spatial structure, and staining, helping emphasize tissue structures over generic image statistics.The authors report improved robustness to resolution and tissue-composition variation, alongside mitigation of mode collapse and structural inconsistency.Source: §7 · p. 11
  • GWDM achieves the lowest FID and KID on BCSS and the lowest KID on PAIP2019, while remaining competitive on MoNuSeg distribution-level metrics.MoNuSeg’s relatively homogeneous, densely packed nuclei leave less scope for differentiation.Source: §7 · p. 11
  • Synthetic augmentation consistently improves mIoU, Dice, and pixel accuracy across DeepLabV3+, UNet++, and SegFormer.The experiments used a self-supervised pretraining and limited-label finetuning pipeline.Source: §7 · p. 11
  • The largest segmentation gains occur on the Hard BCSS subset, while Easy-subset improvements are smaller but consistent.MoNuSeg shows modest, uniform gains across difficulty levels, with baseline performance already near saturation.Source: §7 · p. 12
  • The patch-level implementation and partial reliance on handcrafted descriptors may limit scalability.Future work will explore WSI synthesis and learned pathology embeddings to improve generalisation and reduce manual feature design.Source: §7 · p. 12

8. Conclusion

The conclusion presents GWDM as a wavelet-based, domain-guided diffusion framework for histopathology generation. Across three benchmarks, it reports improved image-generation metrics and segmentation performance with synthetic images.

8. Conclusion

  • GWDM combines domain-specific tissue priors, multiscale morphological, spatial, and staining features, wavelet-based parameterisation, and dual perceptual-structural loss.These components guide reconstruction toward biologically meaningful images with high-frequency cellular detail, stain consistency, and structural fidelity.Source: §8 · p. 12
  • Across BCSS, MoNuSeg, and PAIP2019, GWDM achieves consistent improvements in FID, KID, pFID, and path-FID.The generated images exhibit realistic tissue textures and nuclear organisation aligned with pathological appearance.Source: §8 · p. 12
  • GWDM-synthesised images consistently improve performance across DeepLabV3+, UNet++, and SegFormer segmentation experiments.The conclusion identifies their practical value for training robust segmentation models in clinical digital pathology settings.Source: §8 · p. 12

Related papers

Loading 2610.30775v1…