Privacy-Preserving Synthetic Mammograms: A Generative Model Approach to Privacy-Preserving Breast Imaging Datasets

Accurate synthetic image generation is crucial for addressing data scarcity challenges in medical image classification tasks, particularly in sensor-derived medical imaging. In this work, we propose a novel method using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and nearest-neighbor interpolation to generate high-quality synthetic images for diabetic retinopathy classification. Our approach enhances training datasets by generating realistic retinal images that retain critical pathological features. We evaluated the method across multiple retinal image datasets, including Retinal-Lesions, Fine-Grained Annotated Diabetic Retinopathy (FGADR), Indian Diabetic Retinopathy Image Dataset (IDRiD), and the Kaggle Diabetic Retinopathy dataset. The proposed method outperformed traditional generative models, such as conditional GANs and PathoGAN, achieving the best performance on key metrics: a Fréchet Inception Distance (FID) of 15.21, a Mean Squared Error (MSE) of 0.002025, and a Structural Similarity Index (SSIM) of 0.89 in the Kaggle dataset. Additionally, expert evaluations revealed that only 56.66% of synthetic images could be distinguished from real ones, demonstrating the high fidelity and clinical relevance of the generated data. These results highlight the effectiveness of our approach in improving medical image classification by generating realistic and diverse synthetic datasets.

IntroductionThe aim of this study was to investigate the feasibility of generating synthesized ultrasound biomicroscopy (UBM) images from swept-source anterior segment optical coherent tomography (SS-ASOCT) images using a cycle-consistent generative adversarial network framework (CycleGAN) for iridociliary assessment on a cohort presenting for primary angle-closure screening.MethodsThe CycleGAN architecture was adopted to synthesize high-resolution UBM images trained on the SS-ASOCT dataset from the department of ophthalmology, Xinhua Hospital. The performance of the CycleGAN model was further tested in two separate datasets using synthetic UBM images from two different ASOCT modalities (in-distribution and out-of-distribution). We compared the ability of glaucoma specialists to assess the image quality of real and synthetic images. UBM measurements, including anterior chamber, iridociliary parameters, were compared between real and synthetic UBM images. Intra-class correlation coefficients, coefficients of variation, and Bland–Altman plots were used to assess the level of agreement. The Fréchet Inception Distance (FID) was measured to evaluate the quality of the synthetic images.ResultsThe whole trained dataset included anterior chamber angle images, of which 4037 were obtained by SS-ASOCT and 2206 were obtained by UBM. The image quality of real versus synthetic SS-ASOCT images was similar as assessed by two glaucoma specialists. The Bland–Altman analysis also suggested high consistency between measurements of real and synthetic UBM images. In addition, there was fair to excellent agreement between real and synthetic UBM measurements for the in-distribution dataset (ICC range 0.48–0.97) and the out-of-distribution dataset (ICC range 0.52–0.86). The FID was 21.3 and 24.1 for the synthetic UBM images from the in-distribution and out-of-distribution datasets, respectively.ConclusionWe developed a CycleGAN model to translate UBM images from non-contact SS-ASOCT images. The CycleGAN synthetic UBM images showed fair to excellent reproducibility when compared with real UBM images. Our results suggest that the CycleGAN technique is a promising tool to evaluate the iridociliary and anterior chamber in an alternative non-contact method.

Synthetic medical images have an important role to play in developing data-driven medical image processing systems. Using a relatively small amount of patient data to train generative models that can produce an abundance of additional samples could bridge the gap towards big-data in niche medical domains. These generative models are evaluated in terms of the synthetic data they generate using the Visual Turing Test (VTT), Fréchet Inception Distance (FID), and other metrics. However, these are generally interpreted at the group level, and do not measure the artificiality of individual synthetic images. The present study attempts to address the challenge of automatically identifying artificial images that are obviously-artificial-looking, which may be necessary for filtering out poorly constructed synthetic images that might otherwise deteriorate the performance of assimilating systems. Synthetic computed tomography (CT) images from a progressively-grown generative adversarial network (PGGAN) were evaluated with a VTT and their image embeddings were analyzed for correlation with artificiality. Images categorized as obviously-artificial (≥0. 7 probability of being rated as fake) were classified using a battery of algorithms. The top-performing classifier, a support vector machine, exhibited accuracy of 75.5%, sensitivity of 0.743, and specificity of 0.769. This is an encouraging result that suggests a potential approach for validating synthetic medical image datasets. Clinical Relevance - Next-generation medical AI systems for image processing will utilize synthetic images produced by generative models. This paper presents an approach towards verifying artificial image legibility for quality-control before being deployed for these purposes.

To assess the feasibility of generating synthetic fluorescein angiography (FA) images from color fundus (CF) images using pixel-to-pixel generative adversarial network (pix2pixGANs) for clinical applications. Research questions addressed image realism to retinal specialists and utility for assessing macular edema (ME) in Retinal Vein Occlusion (RVO) eyes. We used a registration-guided pix2pixGANs method trained on the CF-FA dataset from Kham Eye Centre, Kandze Prefecture People's Hospital. A visual Turing test confirmed the realism of synthetic images without novel artifacts. We then assessed the synthetic FA images for assessing ME. Finally, we quantitatively evaluated the synthetic images using Fréchet Inception distance (FID) and structural similarity measures (SSIM). The raw development dataset had 881 image pairs from 349 subjects. Our approach is capable of generating realistic FA images because small vessels are clearly visible and sharp within one optic disc diameter around the macula. Two retinal specialists agreed that more than 85% of synthetic FA images have good or excellent image quality. For ME detection, accuracy was similar for real and synthetic images. FID demonstrated a 38.9% improvement over the previous state-of-the-art (SOTA), and SSIM reached 0.78 compared to the previous SOTA's 0.67. We developed a pix2pixGANs model translating FA images from label-free CF images, yielding reliable synthetic FA images. This suggests potential for noninvasive evaluation of ME in RVO eyes using pix2pix GANs techniques. Pix2pixGANs techniques have the potential to assist in the noninvasive clinical assessment of ME in RVO eyes.

Efficacy of image similarity as a metric for augmenting small dataset retinal image segmentation.

Global Cancer Statistics 2020 states that there are 2.2 million lung cancer cases worldwide with 1.8 million deaths. At present, deep learning based CAD system for lung nodules classification has been extensively explored. However, this approach requires a great size of images which becomes an issue for medical images. Thus, Generative Advesarial Network (GAN) is introduced to ease this limitation by creating synthetic images. In this study, four GAN architectures namely Deep Convolutional (DCGAN), Deep Regret Analytic GAN (DRAGAN), Wasserstein GAN (WGAN) and Wasserstein GAN with Gradient Penalty (WGANGP) are used in generating synthetic medical images which are then used to classify the lung lesions into benign and malignant via ShuffleNet. The classification is assessed based on pecificity, accuracy, sensitivity, and values of AUC-ROC. Experimental results show that DRAGAN achieved the lowest Fréchet Inception Distance (FID) score of 137.48 of the new generated datasets followed by the WGAN-GP (158.86), WGAN (176.86) and DCGAN (172.56). However, due to the lack of diversity in datasets of DRAGAN, instead WGAN-GP ShuffleNet performed the best in the classification task achieving 98.87% of accuracy, 98.36% of specificity, 99.34% of sensitivity and highest AUC among others at 99.96%. Overall, both high quality and well diversed synthetic images are equally important for the lung nodules classification problem.

INTRODUCTION: 5-Aminolevulinic acid (5-ALA) is the most widely used fluorophore in image-guided glioma surgery, and previous work at our institution demonstrated that 5-ALA highlights tumor-associated macrophages (TAMs) in two-photon microscopy images of brain tumor tissue. Using a unique, paired dataset of stimulated Raman histology (SRH) and two-photon images that share one-to-one spatial resolution, we propose a deep-learning approach to identify macrophages from intraoperative SRH images without requiring fluorescent labels. METHODS: We compiled a dataset of 9,554 non-overlapping, 300-by-300 pixel fields of view from paired SRH and two-photon images representing 40 cases of high-grade glioma. A deep generative adversarial network (pix2pix), consisting of a U-Net generator and PatchGAN discriminator, was trained to generate synthetic two-photon images from each SRH patch. The model was trained for 200 epochs, and similarity between the synthetic and real distributions was assessed qualitatively while training and quantitatively using FrÉchet inception distance (FID). RESULTS: Our model was successfully trained to generate synthetic two-photon images nearly indistinguishable from ground truth examples. The FID between a held-out test set of real and synthetic images was 8.58, compared to a mean FID of 9.26 ± 0.24 between randomly sampled sets of real images. Our model consistently inferred the location of brightly fluorescing TAMs in held-out, previously unseen test images. CONCLUSIONS: We used deep learning to visualize TAMs in label-free SRH images. Analysis of TAM infiltration has the potential to identify patients most likely to benefit from immunotherapy clinical trials. Ongoing work will leverage additional deep-learning approaches to automatically identify and quantify TAM infiltration with SRH, enabling rapid patient-specific analysis of the glioma immune microenvironment.

BackgroundMalignant melanoma (MM) is the most aggressive skin cancer, requiring early diagnosis for better outcomes. While deep learning models have shown promise in dermatological image analysis, their performance is constrained by limited training data. Generative Adversarial Networks (GANs) offer a solution by generating synthetic images for data augmentation. However, assessing their clinical reliability remains difficult as automated metrics may not fully capture visual realism or clinical usability.The ObjectiveThis study presents a comprehensive framework for creating high-quality synthetic dermatoscopic pictures of MM lesions, as well as a holistic validation methodology that uses quantitative metrics and qualitative dermatologist assessment to provide a full clinical evaluation of the generated images. Three GAN architectures (DCGAN, StyleGAN2, and StyleGAN3-t) are explored. Lesions on the face, palms, and soles are excluded due to their unique dermoscopic patterns.Materials and MethodsA dataset of 1,774 dermatoscopic body MM images was used to train the models, assessed with Fréchet Inception Distance (FID), Kernel Inception Distance (KID), precision, and recall. Afterwards, a panel of 17 dermatologists with different levels of expertise assessed image quality using a 7-point Likert scale, with accuracy, sensitivity, specificity, and inter-rater agreement analysed.ResultsStyleGAN2 achieved the best image fidelity (FID: 18.89, KID: 0.0025), while StyleGAN3-t demonstrated stable but slower convergence. Both StyleGAN models outperformed DCGAN in diversity and fidelity. The validation study showed that StyleGAN2-generated images were often indistinguishable from real ones, reflected in low specificity and accuracy values among evaluators.ConclusionsThe study highlights the effectiveness of GANs in generating high-quality synthetic images, proposing a validation framework that integrates expert assessments with state-of-the-art quantitative metrics. This approach advances standardisation in GAN evaluation, ensuring synthetic images are clinically relevant for dermatological AI applications.

The use of generative models in image synthesis has become increasingly prevalent. Synthetic medical imaging data is of paramount importance, primarily because medical imaging data is scarce, costly, and encumbered by legal considerations pertaining to patient confidentiality. Synthetic medical images offer a potential answer to these issues. The predominant approaches primarily assess the quality of images and the degree of resemblance between these images and the original ones employed for their generation.The central idea of the work can be summarized in the question: Do the performance metrics of Frechet Inception Distance(FID) and Inception Score(IS) in the Cycle-consistent Generative Adversarial Networks (CycleGAN) model are adequate to determine how real a generated chest x-ray pneumonia image is? In this study, a CycleGAN model was employed to produce artificial images depicting 3 classes of chest x-ray pneumonia images: general(any type), bacterial, and viral pneumonia. The quality of the images were evaluated assessing and contrasting 3 criteria: performance metric of CycleGAN model, clinical assessment of respiratory experts and the results of classification of a visual transformer(ViT). The overall results showed that the evaluation metrics of the CycleGAN are insufficient to establish realism in generated medical images.

In the realm of computational pathology, the scarcity and restricted diversity of genitourinary (GU) tissue datasets pose significant challenges for training robust diagnostic models. This study explores the potential of Generative Adversarial Networks (GANs) to mitigate these limitations by generating high-quality synthetic images of rare or underrepresented GU tissues. We hypothesized that augmenting the training data of computational pathology models with these GAN-generated images, validated through pathologist evaluation and quantitative similarity measures, would significantly enhance model performance in tasks such as tissue classification, segmentation, and disease detection. To test this hypothesis, we employed a GAN model to produce synthetic images of eight different GU tissues. The quality of these images was rigorously assessed using a Relative Inception Score (RIS) of 1.27 ± 0.15 and a Fréchet Inception Distance (FID) that stabilized at 120, metrics that reflect the visual and statistical fidelity of the generated images to real histopathological images. Additionally, the synthetic images received an 80% approval rating from board-certified pathologists, further validating their realism and diagnostic utility. We used an alternative Spatial Heterogeneous Recurrence Quantification Analysis (SHRQA) to assess the quality of prostate tissue. This allowed us to make a comparison between original and synthetic data in the context of features, which were further validated by the pathologist's evaluation. Future work will focus on implementing a deep learning model to evaluate the performance of the augmented datasets in tasks such as tissue classification, segmentation, and disease detection. This will provide a more comprehensive understanding of the utility of GAN-generated synthetic images in enhancing computational pathology workflows. This study not only confirms the feasibility of using GANs for data augmentation in medical image analysis but also highlights the critical role of synthetic data in addressing the challenges of dataset scarcity and imbalance. Future work will focus on refining the generative models to produce even more diverse and complex tissue representations, potentially transforming the landscape of medical diagnostics with AI-driven solutions.

Study question How effectively can a latent diffusion model generate high-fidelity embryo images tailored to the specific contextual needs of researchers, based on the user’s text input? Summary answer The latent diffusion model successfully generated high-resolution embryo images, providing a novel approach to address data scarcity in embryo imaging. What is known already The use of AI in generating synthetic data has been increasingly explored to address the scarcity of real-world datasets in various fields. Particularly in medical imaging, AI-generated synthetic data offers a potential solution to overcome the limitations imposed by data privacy concerns and ethical considerations. Previous studies have shown that AI can replicate complex patterns in data, but its application in generating embryo images remains less explored. Study design, size, duration Single static images of 5,133 Day 5 blastocysts and 2,093 Day 3 cleavages were retrospectively collected from seven in vitro fertilization clinics between June 2011 and May 2022. The images were analyzed along with relevant metadata including clinical information and embryo grades. Day 3 embryo grading was based on the number of blastomeres, evenness, and fragmentation percentage. Day 5 grading criteria included the inner cell mass, trophectoderm, and blastocyst stage evaluations. Participants/materials, setting, methods An AI model using latent diffusion was developed to generate 10,051 Day 5 and 10,088 Day 3 synthetic embryo images. The authenticity was assessed through visual Turing tests, where embryologists discerned real from synthetic images. For the evaluation, 200 real (100 Day 5, 100 Day 3) and 200 synthetic (100 Day 5, 100 Day 3) images were randomly chosen from each dataset, ensuring a comprehensive test of the generated images’ realism. Main results and the role of chance The AI model’s efficacy in generating synthetic embryo images was assessed through visual Turing tests on Day 5 and Day 3 embryos, yielding accuracies of 0.59 and 0.57, respectively. These accuracies fall within the 99% confidence interval of near-random performance (0.41 to 0.59), highlighting the challenge in distinguishing synthetic from real images. For Day 5 embryos, the sensitivity and specificity were recorded at 0.52 and 0.66, indicating a moderate challenge in identifying synthetic images and a relatively higher ease in recognizing real ones. Day 3 embryos presented a lower sensitivity of 0.41, suggesting greater difficulty in detecting synthetic images, while the specificity increased to 0.73, indicating a stronger ability to identify real images. Collectively, with an overall accuracy of 0.58, sensitivity of 0.47, and specificity of 0.70, these findings confirm the synthetic images’ remarkable realism, closely emulating actual embryo features to the extent that differentiation by experts proved challenging. This level of realism underscores the synthetic images’ potential to significantly enrich embryological research datasets, promising advancements in the field. Limitations, reasons for caution The study’s limitation lies in the generated synthetic embryos being based on three embryo grade categories, limiting feature diversity. Additionally, there was no quantitative assessment method for these images, relying instead on expert evaluation, which required deploying a large number of embryologists. Wider implications of the findings The model’s generation of highly realistic embryo images counters data scarcity in embryological research, potentially elevating AI’s utility, enriching educational content, advancing reproductive medicine, and ensuring ethical data usage. Trial registration number not applicable

Medical image translation has become an essential tool in modern radiotherapy, providing complementary information for target delineation and dose calculation. However, current approaches are constrained by their modality-specific nature, requiring separate model training for each pair of imaging modalities. This limitation hinders the efficient deployment of comprehensive multimodal solutions in clinical practice. To develop a unified image translation method using variational autoencoder (VAE) latent space mapping, which enables flexible conversion between different medical imaging modalities to meet clinical demands. We propose a three-stage approach to construct a unified image translation model. Initially, a VAE is trained to learn a shared latent space for various medical images. A stacked bidirectional transformer is subsequently utilized to learn the mapping between different modalities within the latent space under the guidance of the image modality. Finally, the VAE decoder is fine-tuned to improve image quality. Our internal dataset collected paired imaging data from 87 head and neck cases, with each case containing cone beam computed tomography (CBCT), computed tomography (CT), MR T1c, and MR T2W images. The effectiveness of this strategy is quantitatively evaluated on our internal dataset and a public dataset by the mean absolute error (MAE), peak-signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Additionally, the dosimetry characteristics of the synthetic CT images are evaluated, and subjective quality assessments of the synthetic MR images are conducted to determine their clinical value. The VAE with the Kullback‒Leibler (KL)-16 image tokenizer demonstrates superior image reconstruction ability, achieving a Fréchet inception distance (FID) of 4.84, a PSNR of 32.80 dB, and an SSIM of 92.33%. In synthetic CT tasks, the model shows greater accuracy in intramodality translations than in cross-modality translations, as evidenced by an MAE of 21.60±8.80 Hounsfield unit (HU) in the CBCT-to-CT task and 45.23±13.21 HU/47.55±13.88 in the MR T1c/T2w-to-CT tasks. For the cross-contrast MR translation tasks, the results are very close, with mean PSNR and SSIM values of 26.33±1.36 dB and 85.21%±2.21%, respectively, for the T1c-to-T2w translation and 26.03±1.67 dB and 85.73%±2.66%, respectively, for the T2w-to-T1c translation. Dosimetric results indicate that all thegamma pass ratesfor synthetic CTs are higher than99%for photonintensity-modulated radiation therapy (IMRT)planning. However, the subjective quality assessment scores for synthetic MR images are lower than those for real MR images. The proposed three-stage approach successfully develops a unified image translation model that can effectively handle a wide range of medical image translation tasks. This flexibility and effectiveness make it a valuable tool for clinical applications.

Deep Learning (DL) models have outperformed remarkably and effectively on several Computers Vision applications. However, these models require large amounts of data to avoid overfitting problems. Overfitting happens when a network trains a function with an incredibly high variance to represent the training data perfectly. Consequently, medical images lack to availability of large labeled datasets, and the annotation of medical images is expensive and time-consuming for experts, as the COVID-19 virus is an infectious disease, these datasets are scarce and it is difficult to get large datasets. The limited amount of the COVID-19 class compared to any other classes, for example (healthy). To solve the scarcity data problem, we adjust a Conditional Generative Adversarial Network (CGAN) as a solution to the problems of scarcity and limited data. CGAN contains two neural networks: a generator that creates synthetic (fake) images, and a discriminator that recognizes a real sample of training and a generated sample from the generator. The adjusted CGAN is able to Generate synthetic images with high resolution and close to the original images which aid in expanding the limited dataset specific to a new pandemic. In addition to CGAN augmenting strategies, this research also briefly explores additional aspects of data augmentation like time augmentation and total dataset size. Frechet inception distance metric (FID) has been used for evaluating synthetic images generated by CGAN. The adjusted CGAN obtains better FID results for the high-resolution synthetic X-rays images it achieves 2.349%.

Progress in accurate medical image classification is often hampered by concerns surrounding data privacy and scarcity of data for certain medical diseases, leading to sparsity and unbalanced datasets. To address these challenges, this study uses generative models, namely, Denoising Diffusion Probabilistic Models (DDPMs) and Progressive Growing Generative Adversarial Networks (PGGANs), for dataset improvement. In this article, we propose a framework for understanding how the resultant synthetic images generated by DDPM and PGGANs affect four different models’ performance: a specially crafted Convolutional Neural Network, an untrained VGG16, a pretrained VGG16, and a pretrained ResNet50. For modeling practical constraints in real applications, experiments applied Random Sampling and Greedy K Sampling to obtain small unbalanced datasets. Synthetic image quality was also measured by applying Fréchet Inception Distance (FID), and their impact was further explored by comparing classification results with their original datasets. Experiments reveal that DDPM consistently produced images of higher realism, backed by lower FID scores, and overtakes PGGANs in augmenting classification outcomes of all investigated models and datasets. Addition of DDPM-generated images to original datasets obtained improvement of about 6% in accuracy and therefore enhanced robustness and reliability of models, specifically when datasets are unbalanced. Although Random Sampling obtained better consistency, Greedy K Sampling obtained higher variability but higher FID scores. Overall, this research identifies the potential of DDPM to effectively augment and balance sparse datasets of medical images and subsequently improve training of models and predictive outcomes. Received: 1 July 2025 | Revised: 3 September 2025 | Accepted: 15 September 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in GitHub at https://github.com/imankhazrak/DDPM_X-Ray. Author Contribution Statement Iman Khazrak: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project administration. Shakhnoza Takhirova: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation,Writing – original draft, Writing – review & editing, Visualization. Mostafa M. Rezaee: Validation, Investigation, Resources, Writing – review & editing. Mehrdad Yadollahi: Investigation, Resources, Writing – review & editing. Robert C. Green II: Methodology, Writing – review & editing, Supervision. Shuteng Niu: Methodology, Writing – review & editing, Supervision.

Rapid Unpaired CBCT-Based Synthetic CT for CBCT-Guided Adaptive Radiotherapy