Synthetic histopathological image generation for lung adenocarcinoma subtypes: an exploratory diffusion model study

Introduction

Generative artificial intelligence (AI) is a subfield of AI focused on developing models capable of generating data such as images, text, sound, video, or synthetic data (1). These systems use statistical models and machine learning to create new and original content, often indistinguishable from real data (2), and they employ autoregressive transformers with billions of parameters (3,4). Although generative adversarial networks (GANs) have shown promising results in image synthesis, their performance is often limited to datasets with low variability. In contrast, diffusion models, based on noise-removal encoders, have emerged as the leading technology, achieving outstanding results in super-resolution and conditional image synthesis (5), with significant potential for generating realistic, high-quality images (6,7).

Synthetic data, artificially generated through simulations or algorithms, accelerates model training and complements datasets when they are incomplete or difficult to obtain (8). In healthcare, they are key to ensuring fairness, avoiding bias, and preserving privacy by anonymizing sensitive data (9,10). Their value is especially significant when filling gaps in databases or collecting data over time (9). In medicine, the use of synthetic images has been proposed in radiology for computed tomography (CT) scans and X-rays (11), in pathology for breast cancer (12), and in gastroenterology with endoscopic images (13). In histology, the complexity of representing cells, their microenvironment, and organelles, such as the nucleus, is critical, particularly in diagnosing breast cancer (14) and kidney tumors (15). The accuracy and accessibility of this data are essential for effectively training AI models, as they directly impact the models’ sensitivity and effectiveness.

The application of synthetic images in lung cancer research, particularly for adenocarcinoma subtype characterization, is justified by the high incidence and mortality of this disease, as well as by the morphological heterogeneity and uneven representation of certain histological subtypes in curated digital pathology datasets, which continue to challenge the development of robust diagnostic and prognostic AI models. Synthetic images emerge as a valuable solution, enabling the generation of diversified and precise data that can enrich training datasets without the ethical or logistical challenges involved in collecting large volumes of clinical data.

In the present study, we sought to evaluate the technical feasibility of generating synthetic histopathological images of lung adenocarcinoma subtypes using diffusion models fine-tuned with Low-Rank Adaptation (LoRA). Additionally, the generated images were assessed through quantitative similarity metrics, expert validation by pathologists and pathology residents, and downstream evaluation in convolutional neural network classification tasks. The ultimate goal was to explore the potential utility and limitations of synthetic histopathological data as a complementary resource for computational pathology and data augmentation workflows.

Methods

Study type

An exploratory study was conducted to generate and validate synthetic data of lung adenocarcinomas.

Database selection

The LC25000 database was selected due to its size and quality; it includes 25,000 medical images of colon and lung tissues, including 5,000 image patches from lung adenocarcinomas. The dataset was compiled by Andrew Borkowski and his associates at James Hospital Tampa, Florida. All images are de-identified, Health Insurance Portability and Accountability Act (HIPAA)-compliant, validated, and can be freely downloaded by AI researchers; it is hosted on Kaggle with public access (16).

Synthetic image generation

Figure 1 summarizes the complete pipeline used for the generation of synthetic lung adenocarcinoma images using diffusion models based on Stable Diffusion 1.5 fine-tuned with LoRA. Each image patch was rescaled from 768×768 pixels to 512×512 pixels. The generation of synthetic images was performed by fine-tuning the Stable Diffusion model (5) using the LoRA technique (17). The fine-tuning was carried out to adapt the model specifically to the characteristics of lung adenocarcinoma images. Importantly, separate LoRA models were trained for each histological pattern of lung adenocarcinoma (e.g., acinar, micropapillary, solid, and mucinous), allowing the model to learn pattern-specific morphological features.

Figure 1 Methodological workflow for the generation, validation, and downstream evaluation of synthetic histopathological images of lung adenocarcinoma using Stable Diffusion 1.5 fine-tuned with LoRA (figure created by the authors using ChatGPT Imagen 2). AUC, area under the curve; H&E, hematoxylin and eosin; LoRA, Low-Rank Adaptation.

Parameter selection

The parameters used for fine-tuning the Stable Diffusion 1.5 model were selected based on previous reports and empirical recommendations for LoRA training in limited-data environments, defined as follows: UNet learning rate: 0.0005, encoder learning rate: 0.0001, number of epochs: 9–10 (the best of multiple intermediate LoRA checkpoints), batch size: 2, scheduler: cosine with restarts, and scheduler value: 3.

Given that histopathological images of lung cancer represent a visual domain substantially different from the original pretraining space of Stable Diffusion, their learning through LoRA can be considered analogous to acquiring an abstract concept, characterized by high intra-class variability and the absence of a canonical representation (18). Accordingly, a subset of approximately 50–80 representative images per histological pattern was selected for LoRA fine-tuning, consistent with commonly adopted practices for concept-specific adaptation of diffusion models (19).

Development of classification models

To evaluate the impact of synthetic image-based data augmentation on classification performance, two convolutional neural network models were trained using transfer learning with EfficientNet-B0 (20), pretrained on ImageNet. Two experimental conditions were defined: training exclusively with original histopathological images and training with a combined dataset of original and synthetic images generated through LoRA fine-tuning of Stable Diffusion. The test set consisted exclusively of original images and was created using a fixed stratified 20% split, identical across both experimental conditions to ensure comparability of results. The remaining 80% of the original images were further divided into training (85%) and validation (15%) subsets, both stratified by histological subtype. In the augmented condition, all synthetic images were incorporated exclusively into the training set.

Exploratory expert validation of synthetic image realism

As a complementary approach, an exploratory validation by pathologists was conducted to assess the perceptual realism of the generated images. For each set of generated images corresponding to each specific subtype of adenocarcinoma, the most representative cases were selected by consensus among the participating pathologists, ideally 10 for each subtype. Subsequently, 2 to 3 images were randomly selected for each subtype, including both images from the LC25000 database (real controls) and the synthetic images generated (cases). These were used to construct a virtual questionnaire in which participants were asked to determine whether the images were real or synthetic. The questionnaire, consisting of 20 items, was distributed through an open call to pathologists, to which 33 participants responded. This group included 19 practicing pathologists from the Institute of Legal Medicine and the Fundación Universitaria de Ciencias de la Salud (FUCS), as well as 12 pathology residents from FUCS.

Statistical analysis

To quantify the distributional similarity between original and LoRA-generated histopathological images, two complementary metrics were used: Fréchet Inception Distance (FID) (21) and Kernel Inception Distance (KID) (22). High-level image features were extracted using a pretrained InceptionV3 network following a standardized preprocessing pipeline. While FID measures the distance between Gaussian-fitted feature distributions, KID—based on maximum mean discrepancy—provides an unbiased estimate, making it more suitable for small or imbalanced datasets. Both metrics were computed independently for each lung adenocarcinoma subtype.

Model performance was evaluated using macro-averaged one-vs-rest ROC-AUC and macro F1-score, both appropriate for multiclass settings with class imbalance. Ninety-five percent confidence intervals were estimated using percentile bootstrap resampling with 1,000 iterations (23). Statistical significance between the original and augmented training conditions was assessed using a two-sided bootstrap permutation test with 10,000 iterations, considering P<0.05 as significant. All analyses were performed in Python 3.10 using PyTorch 2.x, scikit-learn, and clean-fid.

A descriptive analysis of the participating pathologists was conducted, including level of training, years of professional experience, and performance in distinguishing synthetic from real images. Survey responses were statistically analyzed to evaluate accuracy rates across different participant groups, including practicing pathologists, instructors, and residents. Normality was assessed using the Shapiro-Wilk test. Since several variables did not follow a normal distribution, descriptive statistics were reported as median and interquartile range (IQR). Correlation analyses were performed using Spearman’s rank correlation coefficient, and paired comparisons between recognition rates for synthetic and real images were conducted using the Wilcoxon signed-rank test. Statistical significance was established at P<0.05.

For additional methodological details regarding image selection and labeling (curation), the formal definition of diffusion models, parameter selection for the classification models, and the equipment and software resources used in this study, please refer to Appendix 1. Figure S1 illustrates examples of the annotation process applied to LC25000 images, while Figure S2 shows representative synthetic image outputs generated under different scheduler configurations and training epochs.

Ethical considerations

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics and research committee of the Fundación Universitaria de Ciencias de la Salud (FUCS) University under approval record DI-I-0389-25 dated September 12, 2025. Informed consent was not required as the study did not include any patient data.

Results

From the LC25000 database, we selected 575 high-resolution scanned image patches of lung adenocarcinomas, distributed across the following patterns: enteric (130 images), lepidic (90 images), solid (90 images), mucinous (73 images), colloid (67 images), acinar (59 images), and micropapillary (52 images). For the fetal pattern, only 7 images were found, which made it impossible to use for fine-tuning via LoRA. Models derived from the ninth and tenth epochs were used for image generation. A total of 1,124 synthetic images were generated according to the following patterns: enteric (81 images), lepidic (209 images), solid (169 images), mucinous (256 images), colloid (185 images), acinar (172 images), and micropapillary (52 images).

Table 1 presents the distributional similarity between real and synthetically generated histopathological images. The solid subtype exhibited the lowest FID (75.30) and KID (0.0313) values, suggesting a higher capability of the model to capture its morphological characteristics. In contrast, the mucinous subtype showed the highest FID (151.60) and one of the highest KID values (0.0960), indicating greater difficulty in generating representative images. The remaining subtypes (acinar, colloid, enteric, and lepidic) presented intermediate values, reflecting heterogeneous model performance depending on the histological pattern.

The EfficientNet model trained exclusively with original images achieved a Macro AUC of 0.9385 and a Macro F1-score of 0.7508. Similarly, the model trained with both original and synthetic images obtained a Macro AUC of 0.9389 and a Macro F1-score of 0.7498 (see Table 2). No statistically significant differences were observed between the two approaches (P=0.38), suggesting that synthetic augmentation preserved overall classification performance without significant improvement.

Survey results from pathologists

The curation process led to the final selection of 10 images for the solid pattern, 10 for the acinar pattern, 12 for the micropapillary pattern, and 1 for the mucinous pattern. It was determined that none of the generated images for the enteric, lepidic, and colloid patterns were representative of these tumors’ characteristics. For the evaluation by pathologists, the questionnaire included three randomly selected images from the solid, acinar, and micropapillary patterns, plus one from the mucinous pattern.

A total of 33 surveys were conducted with pathologists, of which two were discarded for not including all responses. Among the participants, 18 (58.1%) were women and 13 (41.9%) were men. Of these, 19 were practicing pathologists, and 12 were residents. The median years of experience for the overall population was 7 years (IQR: 2–16.5 years). Practicing pathologists had a median of 12 years of experience (IQR: 8.5–20 years), whereas residents had a median of 2 years (IQR: 1–2 years).

The median number of correct responses was 5.0 (IQR: 2.0–6.0) for synthetic images and 7.0 (IQR: 6.0–9.0) for real images (Figure 2); a detailed breakdown of the pathologists’ responses for each real and synthetic image evaluated in the survey is provided in Table S1. The Wilcoxon signed-rank test demonstrated a significant difference between recognition performance for synthetic and real images (P=0.002), with higher accuracy for real images. No significant correlations were identified between years of experience and performance for either synthetic image recognition (Spearman’s rho =−0.201, P=0.28) or real image recognition (Spearman’s rho =0.158, P=0.39) (Table 3). Overall, pathologists performed significantly better in identifying real images compared to synthetic images, without evidence that years of experience substantially influenced performance.

Figure 2 Comparison of correct responses in synthetic and real images by groups (source: the authors). IQR, interquartile range.

Discussion

Overall, our findings are consistent with previous studies on diffusion models in digital pathology, which suggest that synthetic images can closely resemble real histopathological data, although performance may vary according to tissue complexity. These studies emphasize that evaluating synthetic histopathological images requires a multifaceted approach combining quantitative metrics, downstream model performance, and expert validation (24). The solid subtype showed the best performance, likely due to its more homogeneous morphology, whereas the mucinous subtype performed worst, probably because of its greater morphological variability and dataset imbalance. The concordance between FID and KID rankings further supports the robustness and methodological consistency of these findings.

In our study, although many synthetic images were generated, expert pathologists considered only a limited subset to be sufficiently representative of the histological subtypes for inclusion in the evaluation process, as many generated images still showed limitations in realism and failed to reproduce key histopathological features recognizable by experienced observers. However, despite these perceptual limitations, the incorporation of synthetic images into the convolutional neural network (CNN) training process did not substantially deteriorate overall model performance and achieved results comparable to training with original images alone. These findings suggest that human perceptual realism and downstream computational utility are not necessarily equivalent, as synthetic images that remain identifiable to expert observers may still preserve latent discriminative features exploitable by deep learning models.

Previous studies have demonstrated the potential of synthetic data generation in digital pathology using GAN- and diffusion-based approaches. Chen et al. showed that GAN-generated synthetic renal carcinoma images improved CNN classification performance when combined with real data, although they also highlighted the challenge of representing rare subtypes such as chromophobe carcinoma (9). Similarly, the SNOW dataset, generated with StyleGAN2 and weakly supervised annotation, achieved competitive nuclei segmentation performance in breast cancer and was validated using both FID metrics and expert pathologist evaluation, supporting the utility of synthetic data for pretraining in digital pathology tasks (11).

More recently, diffusion-based approaches have shown promising results in generating realistic histopathological images. Italian researchers generated synthetic images of multiple tissue types using diffusion models trained on GTEx data, achieving image quality and diversity comparable to real samples according to FID and expert assessment (24). In parallel, the RNA-GAN model developed by Carrillo et al. integrated RNA expression data with histopathological images, generating realistic lung and cortical tissue images more efficiently than conventional GANs and highlighting the potential of synthetic data to address data scarcity in computational pathology (25).

The evaluation of synthetic image generation is complex, as it requires measuring not only visual quality but also fidelity and consistency with the initial specifications. While automated metrics such as FID are commonly used to compare the distance between feature distributions of real and generated images (21), these have been criticized for not adequately capturing the diversity of generated content and for their sensitivity to sample size. Although more robust metrics are being developed, such as CLIP Maximum Mean Discrepancy (CMMD), which uses Contrastive Language-Image Pre-training (CLIP) embeddings and the maximum mean discrepancy distance (26), human evaluation remains crucial, as it can detect nuances and errors that automated metrics often overlook, though it introduces challenges due to subjectivity and variability among evaluators.

An important finding was the discrepancy between human perceptual evaluation and downstream CNN performance. Although pathologists identified substantial limitations in realism, the synthetic images still preserved latent discriminative patterns exploitable by deep learning models. Previous studies have suggested that the perceptual quality and histopathological realism of synthetic images, particularly when assessed by expert observers, may partially predict their usefulness for downstream data augmentation tasks, as unrealistic or artifact-prone images can fail to provide meaningful discriminative information for classification models (24,27). Similarly, Niehues et al. demonstrated that synthetic histopathological images with better expert ratings and more favorable quantitative metrics were associated with improved downstream classification performance (28).

The efficiency of synthetic image generation workflows remains a major challenge in medical applications, as large numbers of images often need to be generated to obtain only a small subset with sufficient clinical realism and utility. This limitation is closely related to the need for highly specialized annotation and expert-driven fine-tuning, since training images must accurately reflect true pathological features. In addition, the limited adaptation of general-purpose diffusion models to the medical domain frequently results in unrealistic or clinically unrepresentative images, making the workflow labor-intensive, dependent on expert curation, and overall inefficient due to the high proportion of discarded images.

A limitation of this study is that the CNN models were developed using a relatively small number of cases, which may limit the generalizability of the results. However, this component was intended as an exploratory analysis to assess whether synthetic images could preserve discriminative features useful for deep learning classification tasks. Future work should focus on improving the realism and efficiency of diffusion-based synthetic image generation, expanding validation with larger multicenter datasets, and evaluating the impact of synthetic data on more robust deep learning architectures and clinically relevant diagnostic tasks.

Conclusions

This study demonstrates the technical proof-of-feasibility of generating synthetic histopathological images of lung adenocarcinoma using LoRA-based diffusion models trained on relatively small datasets. Quantitative evaluation revealed heterogeneous performance across histological subtypes, with more morphologically homogeneous patterns achieving greater similarity to real images. Although expert pathologists identified important limitations in realism and considered only a small subset of generated images sufficiently representative, the incorporation of synthetic images into CNN training did not significantly deteriorate classification performance, suggesting that synthetic data may still preserve discriminative features useful for deep learning models. Nevertheless, the current workflow remains labor-intensive and inefficient due to the need for extensive expert annotation and manual curation.

Acknowledgments

During the preparation of this work, the authors used ChatGPT (OpenAI) to assist with language refinement, manuscript editing for clarity and conciseness, support in code development, and the generation of illustrative images. All scientific content, analyses, interpretations, and final editorial decisions were reviewed and approved by the authors, who take full responsibility for the content of the publication.

Footnote

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0056/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0056/prf

Funding: This study is part of the GLORIA project, which was supported by Minciencias under contract number CT 789-2023.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0056/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics and research committee of the Fundación Universitaria de Ciencias de la Salud (FUCS) University under approval record DI-I-0389-25 dated September 12, 2025. Informed consent was not required as the study did not include any patient data.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Kanbach DK, Heiduk L, Blueher G, et al. The GenAI is out of the bottle: generative artificial intelligence from a business model innovation perspective. Rev Manag Sci 2024;18:1189-220.
2. Nikolenko SI. Generative Models in Deep Learning. In: Synthetic Data for Deep Learning. Springer Optimization and Its Applications, vol 174. Cham: Springer; 2021.
3. Razavi A, van den Oord A, Vinyals O. Generating Diverse High-Fidelity Images with VQ-VAE-2. In: Advances in Neural Information Processing Systems. Curran Associates, Inc.; 2019. [cited april 26, 2024]. Available online: https://papers.nips.cc/paper_files/paper/2019/hash/5f8e2fa1718d1bbcadf1cd9c7a54fb8c-Abstract.html
4. Ramesh A, Pavlov M, Goh G, et al. Zero-Shot Text-to-Image Generation. In: Proceedings of the 38th International Conference on Machine Learning. PMLR; 2021:8821-31. [cited april 26, 2024]. Available online: https://proceedings.mlr.press/v139/ramesh21a.html
5. Rombach R, Blattmann A, Lorenz D, et al. High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2022:10684-95. doi: 10.1109/CVPR52688.2022.01042.
6. Dhariwal P, Nichol A. Diffusion Models Beat GANs on Image Synthesis. In: Advances in Neural Information Processing Systems. Curran Associates, Inc.; 2021: 8780-94. [cited april 26, 2024]. Available online: https://papers.nips.cc/paper/2021/hash/49ad23d1ec9fa4bd8d77d02681df5cfa-Abstract.html
7. Ho J, Saharia C, Chan W, et al. Cascaded Diffusion Models for High Fidelity Image Generation. Journal of Machine Learning Research 2022;23:1-33.
8. Abufadda M, Mansour K. A Survey of Synthetic Data Generation for Machine Learning. In: 22nd International Arab Conference on Information Technology (ACIT). Muscat, Oman; 2021.
9. Chen RJ, Lu MY, Chen TY, et al. Synthetic data in machine learning for medicine and healthcare. Nat Biomed Eng 2021;5:493-7. [Crossref] [PubMed]
10. Gao C, Killeen BD, Hu Y, et al. Synthetic data accelerates the development of generalizable learning-based algorithms for X-ray image analysis. Nat Mach Intell 2023;5:294-308. [Crossref] [PubMed]
11. Ding K, Zhou M, Wang H, et al. A Large-scale Synthetic Pathological Dataset for Deep Learning-enabled Segmentation of Breast Cancer. Sci Data 2023;10:231. [Crossref] [PubMed]
12. Vats A, Pedersen M, Mohammed A, et al. Evaluating clinical diversity and plausibility of synthetic capsule endoscopic images. Sci Rep 2023;13:10857. [Crossref] [PubMed]
13. Mahmood F, Borders D, Chen RJ, et al. Deep Adversarial Training for Multi-Organ Nuclei Segmentation in Histopathology Images. IEEE Trans Med Imaging 2020;39:3257-67. [Crossref] [PubMed]
14. Hossain MS, Sakib N. Renal Cell Cancer Nuclei Segmentation from Histopathology Image Using Synthetic Data. 2020 16th IEEE International Colloquium on Signal Processing & Its Applications (CSPA), Langkawi, Malaysia, 2020:236-241. doi: 10.1109/CSPA48992.2020.9068701.
15. Kolárik M, Gojdičová L, Paralič J. Principles of Synthesizing Medical Datasets. Acta Electrotechnica et Informatica 2022;22:25-9.
16. Borkowski A. LC25000 [dataset]. Zenodo; 2025. Available online: https://zenodo.org/records/14998042
17. Hu EJ, Shen Y, Wallis P, et al. LoRA: Low-Rank Adaptation of Large Language Models. 2021. [cited april 26, 2024]. Available online: https://openreview.net/forum?id=nZeVKeeFYf9
18. Osorio P, Jimenez-Perez G, Montalt-Tordera J, et al. Latent Diffusion Models with Image-Derived Annotations for Enhanced AI-Assisted Cancer Diagnosis in Histopathology. Diagnostics (Basel) 2024;14:1442. [Crossref] [PubMed]
19. Ruiz N, Li Y, Jampani V, et al. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023;22500-10.
20. Tan M, Le Q. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. En: Proceedings of the 36th International Conference on Machine Learning. PMLR; 2019:6105-14. [cited may 6, 2026]. Available online: https://proceedings.mlr.press/v97/tan19a.html
21. Heusel M, Ramsauer H, Unterthiner T, et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. In: Advances in Neural Information Processing Systems. Curran Associates, Inc.; 2017. [cited april 26, 2024]. Available online: https://papers.nips.cc/paper_files/paper/2017/hash/8a1d694707eb0fefe65871369074926d-Abstract.html
22. Binkowski M, Sutherland DJ, Arbel M, et al. In: The Sixth International Conference on Learning Representations - ICLR 2018. Vancouver Convention Center, Vancouver CANADA; 2018. [Cited July 1, 2026]. Available online: https://www.semanticscholar.org/paper/Demystifying-MMD-GANs-Binkowski-Sutherland/9723066a5587e6267d8abfd7feefd0637a5a211c
23. Efron B, Tibshirani RJ. An Introduction to the Bootstrap. New York: Chapman and Hall/CRC; 1994:456.
24. Pozzi M, Noei S, Robbi E, et al. Generating and evaluating synthetic data in digital pathology through diffusion models. Sci Rep 2024;14:28435. [Crossref] [PubMed]
25. Carrillo-Perez F, Pizurica M, Ozawa MG, et al. Synthetic whole-slide image tile generation with gene expression profile-infused deep generative models. Cell Rep Methods 2023;3:100534. [Crossref] [PubMed]
26. Jayasumana S, Ramalingam S, Veit A, et al. Rethinking FID: Towards a Better Evaluation Metric for Image Generation. En: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2024:9307-15. [Cited July 1, 2026]. doi: 10.1109/CVPR52733.2024.00889.
27. Benito-Del-Valle L, Alvarez-Gila A, Eguskiza I, et al. Unleashing the Potential of Synthetic Images: A Study on Histopathology Image Classification. In: Del Bue A, Canton C, Pont-Tuset J, et al. Computer Vision – ECCV 2024 Workshops. Cham: Springer Nature Switzerland; 2025:139-55.
28. Niehues JM, Müller-Franzes G, Schirris Y, et al. Using histopathology latent diffusion models as privacy-preserving dataset augmenters improves downstream classification performance. Comput Biol Med 2024;175:108410. [Crossref] [PubMed]