AI-Generated Data Improves Multiple Sclerosis Classification in a Basal Ganglia Radiomics Model

AI-Generated Data Improves Multiple Sclerosis Classification in a Basal Ganglia Radiomics Model | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article AI-Generated Data Improves Multiple Sclerosis Classification in a Basal Ganglia Radiomics Model Erick Eduardo López-Ríos, Francisco J. Alvarez-Padilla This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6798846/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 11 Nov, 2025 Read the published version in BMC Medical Imaging → Version 1 posted 16 You are reading this latest preprint version Abstract Background: The heterogeneity and limited availability of magnetic resonance imaging (MRI) datasets in multiple sclerosis (MS) restrict the robustness of quantitative analysis methods like radiomics. This study explores the use of Generative Adversarial Networks (GANs) as a data harmonization technique. This study assesses whether GAN-generated images are realistic enough to improve the performance of radiomics-based classification models. Methods: We trained GANs to synthesize realistic T1w MRI from a cohort of MS patients and healthy controls. Segmented real and GAN-generated images were processed to extract statistics, texture, and shape radiomic features. Different machine-learning classifiers were trained using traditional augmentation techniques and cGAN-augmented datasets to assess pertinence. Explainable AI methods (SHAP) identified the most influential radiomic biomarkers and how they behave between real and GAN datasets. Results: GAN-generated images increased the mean classification accuracy when using a ResNet (from 0.88 to 0.98) on unseen test data. Explainability analyses revealed that texture heterogeneity and specific shape descriptors of the basal ganglia were the top predictors distinguishing MS from controls. Both datasets features show the same behavior when comparing their distributions. Conclusion: Integrating AI-powered synthetic MRI data into radiomic pipelines substantially improves disease classification accuracy and robustness. This approach addresses data scarcity due to differences in scan protocols, uncovers imaging biomarkers, and offers a scalable strategy for enhancing clinical decision support in MS diagnostics. Generative AI Feature Engineering Computer-Aided Diagnosis Multiple Sclerosis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Multiple sclerosis (MS) is a chronic neurological disease that affects the central nervous system [1–3]. Magnetic resonance imaging (MRI) is crucial for diagnosing MS, monitoring its progression, and assessing treatment response [4–5]. Recently, subcortical structures, such as the basal ganglia, are getting recognized for their relevance in MS pathogenesis and their association with diverse clinical manifestations [6–11]. Quantitative MRI analysis, leveraging radiomics, offers a powerful approach to extract high-throughput information. This information holds significant potential for developing machine learning models aimed at MS classification and prognosis [12–14]. Despite its great potential, its clinical applicability and reproducibility are compromised by the quality and nature of the underlying data. Clinical datasets are inherently heterogeneous, as they are often collected from multiple centers using different scanner manufacturers, acquisition protocols, and parameter settings. This difference introduces systematic biases and noise into the images, rendering radiomic features unstable and difficult to reproduce across studies. To increase the datasets and mitigate the few samples that have been acquired with different configurations, traditional data augmentation techniques, such as rotation, flipping, scaling, and intensity adjustments have been used. However, when applied to radiomic extraction, these methods can inadvertently modify the quantitative features of the images. Even slight changes may disrupt texture statistics and histogram distributions, leading to discrepancies that fail to accurately reflect the underlying anatomical structures. These issues are exemplified in Fig. 1 . In this context, the generation of synthetic data using GANs has emerged as a promising data augmentation strategy in medical imaging [15–17]. cGANs possess the capacity to generate realistic images conditioned by specific features or classes, rendering them particularly well-suited for augmenting medical datasets. However, the application of cGANs in this domain still faces open questions concerning the fidelity of synthetic data in reflecting the full spectrum of clinical variations and the optimal methods for validating their utility in downstream analytical tasks [18]. In this study, we explore the use of cGANs to synthesize T1-weighted MRI images from both MS patients and healthy controls (HC) drawn from a multi-center cohort, aiming to increase images acquired with different scanners and protocols, making the model more likely to generalize in new cohorts. Our primary objective is to assess if augmenting the training dataset with these generated images can improve the performance of a classification model based on radiomic features. These features are extracted from the segmented structures, basal ganglia, white and gray matter. We hypothesize that increasing the data variability through synthetic augmentation will mitigate the limitations imposed by small sample sizes and enhance classification accuracy. This evaluation seeks to elucidate both the strengths and limitations of this data augmentation technique. 2 Materials and Methods 2.1 Data Description To carry out this study, a retrospective dataset composed of T1-weighted MRI scans drawn from different mexican radiology centers was used. This dataset included a total of 166 T1-weighted MRI scans, corresponding to 144 distinct patients diagnosed with multiple sclerosis, as well as 175 scans from healthy subjects or controls. Prior to analysis, all images underwent a comprehensive preprocessing procedure. This preprocessing encompassed skull stripping, voxel spacing normalization to dimensions of 1x1x1mm, image reorientation to the RPS (Right-Posterior-Superior) coordinate system, volume cropping or padding to homogenize dimensions to 128x128x128 voxels, and finally, signal intensity normalization. This step was performed differently depending on the subsequent use of the images: for image generation training, intensities were normalized to the range − 1 to 1; for segmentation to the range 0 to 1, and for radiomic extraction scaled using a Z-score normalization. 2.2 Segmentation Atlas for Radiomic Extraction We constructed a high-resolution segmentation atlas to delineate the regions of interest (ROIs) for downstream radiomic analysis, including bilateral thalamus, putamen, globus pallidus, caudate nucleus, as well as whole white and gray matter. All ROIs were initially segmented using a hybrid pipeline that combines a region‐growing algorithm with a geodesic distance‐based support vector machine (SVM) classifier [19], yielding candidate masks. Neuroanatomists subsequently performed manual refinements in 3D-Slicer [20] to ensure anatomical accuracy and consistency. Inter‐rater reliability, assessed on a subset of 20 volumes, yielded a Dice similarity coefficient of 0.92 ± 0.03 for subcortical nuclei and 0.95 ± 0.02 for white/gray matter, confirming reproducibility. An example of the resulting segmentations is shown in Fig. 2 . To streamline future processing, particularly for the cGAN‑generated MRI volumes, a 3D UNet [21] was then trained on these masks, enabling fully automated segmentation of synthetic images with minimal manual intervention. 2.3 Traditional Data Augmentation To compare with data generated from GANs, we applied traditional data augmentation techniques to our input images. Images were randomly translated along a random axis with displacements ranging from − 10 to 10 pixels and rotated by angles between − 15° and 15°. Additionally, random zooming was applied to simulate variations in object scale, while Gaussian noise was introduced to mimic sensor noise. Horizontal mirroring was also performed by flipping the images along the x-axis. 2.4 Image Generation To address the limitation of data availability and investigate the potential of data augmentation, a cGAN was implemented to generate synthetic T1-weighted MRI for both MS patients and HC. The conditional aspect of the GAN was achieved by conditioning the network on the corresponding class label (MS or HC), allowing controlled generation of images specific to each group. The generator architecture, based on the UNet framework [21], and the PatchGAN-based discriminator [22] are illustrated in Fig. 3 . After training the cGAN, 100 novel synthetic images per condition were generated, creating a synthetic dataset of comparable size and balanced classes. Finally, the generated synthetic images were used as input to a 3D UNet network for basal ganglia and white/gray matter segmentation, enabling the extraction of radiomic features. 2.5 Radiomic Features The extraction of radiomic features was performed using the PyRadiomics library [23] and followed the recommendations from the Image Biomarker Standardization Initiative (IBSI) [24]. The feature sets calculated included: Gray Level Co‑occurrence Matrix (GLCM); Gray Level Dependence Matrix (GLDM); Gray Level Run Length Matrix (GLRLM); Gray Level Size Zone Matrix (GLSZM); Shape‑based; and Statistical features. For each segmented basal ganglia structure, a total of 107 radiomic features were computed, encompassing 14 shape‑based features, 18 histogram features, 24 GLCM, 14 GLDM, 16 GLRLM, 16 GLSZM, and 5 NGTDM features (total of 1070 features per subject). To mitigate redundancy and prevent overfitting in downstream modeling, we applied a supervised feature–selection pipeline based on LASSO (Least Absolute Shrinkage and Selection Operator) regression with 10‑fold cross‑validation. This process yielded a final set of 96 radiomic features. A detailed list of all the remaining features is provided in Table A2. These LASSO‑selected features were then used to train machine learning classifiers, SVM, Random Forest, and a Residual Neural Network (ResNet), on the task of distinguishing MS patients from HC. Classifier performance was evaluated in terms of accuracy, F1‑Score, sensitivity, and specificity on an independent test set. To interpret the contribution of each radiomic descriptor to the classifier decisions, we employed SHAP (SHapley Additive exPlanations) values [25]. SHAP provides a unified measure of feature importance by attributing to each feature the change in the model’s output when conditioning on that feature’s presence. We computed SHAP values for the best‑performing model on the test cohort, generating global importance rankings of the selected radiomic features as well as individualized explanations for each subject. 2.6 Classification Algorithms In this study, we employ a diverse set of three classification algorithms, Support Vector Machine (SVM), Random Forest (RF) and ResNet, to analyze radiomic features and to identify the most relevant predictors. For each classifier, experiments were conducted using three training configurations. In the first configuration, models were trained solely on the real dataset, while in the second, the training data were augmented with GAN images generated via cGANs. The last configuration involved the traditional data augmentation instead of the GAN-based. Note that the validation and testing sets were always composed entirely of real data to ensure an unbiased evaluation of generalization performance. For classifiers that do not require an explicit validation phase, we partitioned the real data into 70% for training and 30% for testing. However, for the ResNet a 70%/15%/15% split was used for training, validation, and testing, respectively. All models were rigorously evaluated using multiple metrics, including accuracy, F1-score, sensitivity and specificity. The SVM classifier identifies an optimal hyperplane to maximize the margin between classes and is well suited to capturing complex nonlinear patterns [26]. Its performance was optimized by tuning the kernel function, the penalty parameter C , and the kernel coefficient γ via grid search with 5-fold cross-validation. The RF classifier is an ensemble method that builds and aggregates multiple decision trees, was similarly tuned by varying the number of trees and the maximum tree depth [27]. Finally, the ResNet leverages multiple residual layers of interconnected neurons to capture intricate patterns in the data [28], and its training incorporated early stopping based on validation performance. Selected hyperparameters for each classifier can be consulted in Table A1. 2.7 Implementation Details All experiments were performed on a workstation featuring an NVIDIA L4 GPU (22.5GB VRAM, 20.7GB utilized during training) and 53GB of RAM (approximately 19.29GB in use). Python (latest version at the time) and TensorFlow (latest version) were used, along with libraries such as Nibabel, NumPy, SciPy, scikit-learn, and PyRadiomics. For the cGAN training both the generator and discriminator networks were trained using the Adam optimizer, with an initial learning rate of 0.0002 and a batch size of 4. The generator loss function was a combination of adversarial loss and L1 loss, while the discriminator loss function was binary cross-entropy loss. The output activation function for the generator network was tanh, and for the discriminator network, was sigmoid. The GAN was trained for 200 epochs. Activation functions for the generator and discriminator networks were LeakyReLU and ReLU, respectively. Figure 4 illustrates the complete pipeline of this work. 3 Results 3.1 Image Generation Results The quality of the GAN T1-weighted MRI images generated by the cGAN was quantitatively evaluated using the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) metrics. The results revealed that the GAN images exhibit considerable structural and perceptual similarity to the real images, achieving average values of 0.9026 for SSIM and 33.82dB for PSNR. Figure 5 presents examples of the synthetic MRI images generated by the cGAN compared to real and transformed with data augmentation samples. 3.2 Classification Performance In this study, two experiments were conducted to evaluate the performance of different classification algorithms. In Experiment 1, all radiomic features extracted from differentiated tissue regions, prior to any feature-selection procedure, were used to train the classification algorithms. In Experiment 2, only the subset of radiomic features selected via the LASSO test was used for training. Classifiers were trained using exclusively real data as well as a combination of real and traditionally augmented (TDA) or generative based (GDA) data (applied only in the training set). Tables 1 and 2 summarize the performance of the classifiers measured on the test set using the metrics of sensitivity, specificity, F1‐Score, and accuracy. Table 1 Experiment 1: Performance of classifiers using all radiomic features extracted from differentiated tissues (before LASSO). Classifier Dataset F1-Score Accuracy Sensitivity Specificity SVM Real 0.8245 0.7826 0.6506 0.9886 SVM Real + GDA 0.8479 0.8289 0.7590 0.9318 SVM Real + TDA 0.6116 0.6969 0.92 0.3207 RF Real 0.9532 0.9529 0.9759 0.9318 RF Real + GDA 0.9473 0.9479 0.9879 0.9090 RF Real + TDA 0.9122 0.9162 0.9879 0.8409 ResNet Real 0.9473 0.9461 0.9518 0.9431 ResNet Real + GDA 0.9590 0.9580 0.9638 0.9545 ResNet Real + TDA 0.8771 0.8826 0.9518 0.8068 3.3 Model Interpretability To deepen the decision mechanisms of our classification models, we used the XAI SHAP technique to assess the global importance of radiomic features, obtained from the SHAP values calculated on the ResNet after LASSO selection, both with real and synthetic data. This analysis generates the violin plots in Fig. 6 , where the y-axis features reflect the contribution to the model output, allowing us to identify which ones play a more decisive role in discriminating between MS patients (positive class) and healthy controls (negative class). Table 2 Experiment 2: Performance of classifiers using only radiomic features selected by LASSO. Classifier Dataset F1-Score Accuracy Sensitivity Specificity SVM Real 0.7454 0.5882 0.4761 0.9117 SVM Real + GDA 0.7818 0.625 0.4761 0.9705 SVM Real + TDA 0.8 0.6666 0.5238 0.9705 RF Real 0.8909 0.85 0.8095 0.9411 RF Real + GDA 0.9090 0.8648 0.8619 0.98 RF Real + TDA 0.9272 0.9 0.9571 0.9705 ResNet Real 0.9082 0.8837 0.9523 0.8970 ResNet Real + GDA 0.9541 0.9382 0.9268 0.9705 ResNet Real + TDA 0.9908 0.9879 0.9898 0.9852 The violins illustrate the density of SHAP values, revealing in which ranges most of the observations for each feature are concentrated. Blue indicates a smaller impact on classification, while red indicates a more pronounced effect. Likewise, SHAP values greater than zero (> 0) correspond to features that favor classification in the positive class (EM). Violin plots collect SHAP values from two previously defined scenarios: (a) exclusively on the radiomic features in the real set, and (b) on the extended set with synthetic samples generated by GAN. Although each experiment provides different nuances in the significance distribution, we observe that several key features are repeated in both cases, which reinforces their predictive relevance. Prominent among these are Elongation Putamen, Surface Volume Ratio Gray Matter and Sphericity White Matter, whose SHAP distributions not only show a consistent impact on MS patient classification but also suggest that the inclusion of synthetic data does not drastically alter their influence. 4 Discussion When synthetic images were incorporated into training, all classifiers showed gains over models trained on real data alone or with TDA. In Experiment 1 the ResNet accuracy rose from 94.61–95.80% with GDA. Unlike simple geometric transforms, GAN-based augmentation appears to expose the classifier to novel radiomic heterogeneity, reducing overfitting and improving generalization. In Experiment 2, the ResNet trained on LASSO-selected features reached a F1-Score of 0.9908 with GDA, compared to 0.9382 with TDA and 0.8837 on real data alone. This underscores that adding synthetic samples not only boosts classification metrics but also sharpens the signal of the most discriminative radiomic descriptors. The SHAP analysis offered further insights into the performance. Across both real-only and GAN-augmented datasets, features such as Elongation of the Putamen, Surface-to-Volume Ratio of Gray Matter, and Sphericity of White Matter consistently ranked highest in their contributions to MS classification. Importantly, the inclusion of synthetic data did not drastically reshuffle feature importance, but rather reinforced the existing hierarchy, suggesting that GAN-generated scans preserve the key radiomic signatures of disease. Nevertheless, some limitations exist. First, our current work focused on cross-sectional data, whether models trained with GAN-augmented data can accurately predict future progression remains to be tested. Second, although feature stability across real and synthetic sets is promising, further testing is needed to rule out subtle biases, for instance by analyzing SHAP interactions or conducting domain adaptation studies. Looking ahead, integrating multimodal GANs could further enrich virtual cohorts and enhance robustness across imaging protocols. Finally, pairing synthetic augmentation with active learning strategies could optimize when and how to request additional real scans, maximizing clinical impact while minimizing acquisition costs. 5 Conclusion We have presented a flexible pipeline that combines radiomic feature extraction with GAN‑augmented MRI data to mitigate the challenges of limited and homogeneous datasets. Through comprehensive SHAP analyses, we demonstrated that a subset of texture and shape descriptors from basal ganglia structures provides consistent, interpretable insights into tissue alterations, insights that may reflect neurodegenerative changes but are not specific to multiple sclerosis alone. Rather than predicting individual disease onset, our framework highlights radiomic markers whose longitudinal shifts could guide further clinical investigation or triage for advanced imaging and biomarker testing. Crucially, these models are designed to complement, not replace clinical expertise, serving as an initial step toward integrating imaging‑based signals with diverse clinical, genetic, and biochemical data. Finally, our results confirm that GAN‑generated images can faithfully expand real‑world variability, bolster the robustness of downstream classifiers and offer a reusable resource for future radiomic and deep‑learning studies. Declarations Ethics approval and consent to participate This study was a retrospective analysis of fully anonymized and de-identified medical images. The Hospital Civil de Guadalajara Committee determined that this research did not constitute human subjects research and waived the requirement for formal ethics committee approval and individual informed consent. The study was conducted in accordance with the principles of the Declaration of Helsinki. Consent for publication Not applicable. This study does not include any identifiable personal data or images. Availability of data and materials The data presented in this study are not publicly available due to internal privacy policies regarding their collection and use. However, the dataset can be made available upon reasonable request to the corresponding author, provided that the requesting party complies with the ethical and privacy considerations established by the collaborating institutions. Competing interests The authors declare no competing interests. Funding This research received no external funding. Authors’ contributions This research was conceptualized by EE and FJ. EE led the design and implementation of the methodology, carried out the investigation, including data curation, software development, and formal analysis. EE also prepared the original draft of the manuscript and created the visualizations. FJ supervised the entire research process, guided the theoretical and methodological framing, and contributed significantly to the critical revision and refinement of the manuscript. Both authors reviewed and approved the final version of the manuscript. References Wang L, et al. Human autoimmune diseases: a comprehensive update. Journal of Internal Medicine. 2015;278(4):369–395. Kalinin I, et al. The impact of intracortical lesions on volumes of subcortical structures in multiple sclerosis. American Journal of Neuroradiology. 2020;41(5):804–808. Dobson R and Giovannoni G. Multiple sclerosis–a review. European Journal of Neurology. 2019;26(1):27–40. McGinley MP, et al. Diagnosis and treatment of multiple sclerosis: a review. Jama. 2021;325(8):765–779. Murray T. Diagnosis and treatment of multiple sclerosis. BMJ. 2006;332(7540):525–527. Jakimovski D, et al. Long-standing multiple sclerosis neurodegeneration: volumetric magnetic resonance imaging comparison to Parkinson’s disease, mild cognitive impairment, Alzheimer’s disease, and elderly healthy controls. Neurobiology of Aging. 2020;90:84–92. Schoonheim MM, et al. Subcortical atrophy and cognition: sex effects in multiple sclerosis. Neurology. 2012;79(17):1754–1761. Azevedo CJ, et al. Thalamic atrophy in multiple sclerosis: a magnetic resonance imaging marker of neurodegeneration throughout disease. Annals of neurology. 2018;83(2):223–234. Zivadinov R, et al. Thalamic atrophy is associated with development of clinically definite multiple sclerosis. Radiology. 2013;268(3):831–841. Schoonheim MM, et al. Disability in multiple sclerosis is related to thalamic connectivity and cortical network atrophy. Multiple Sclerosis Journal. 2022;28(1):61–70. Trufanov A, et al. Basal ganglia atrophy as a marker of multiple sclerosis progression. Biomarkers in Neuropsychiatry. 2023;9:100073. Mayerhoefer ME, et al. Introduction to radiomics. Journal of Nuclear Medicine. 2020;61(4):488–495. Tavakoli H, et al. Investigating the ability of radiomics features for diagnosis of the active plaque of multiple sclerosis patients. Journal of Biomedical Physics & Engineering. 2023;13(5):421. Luo X, et al. Multi-lesion radiomics model for discrimination of relapsing-remitting multiple sclerosis and neuropsychiatric systemic lupus erythematosus. European Radiology. 2022;32(8):5700–5710. Barile B, et al. Data augmentation using generative adversarial neural networks on brain structural connectivity in multiple sclerosis. Computer Methods and Programs in Biomedicine. 2021;206:106113. Brugnara G, et al. Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis. Radiology: artificial intelligence. 2024;6(6):e230514. Suganthi K, et al. Review of medical image synthesis using GAN techniques. In: ITM Web of Conferences. vol. 37. EDP Sciences; 2021. p. 01005. de Farias EC, et al. Impact of GAN-based lesion-focused medical image super-resolution on the robustness of radiomic features. Scientific Reports. 2021;11(1):21361. Park S, et al. Seed growing for interactive image segmentation using SVM classification with geodesic distance. Electronics Letters. 2017;53(1):22–24. Pieper S, et al. 3D Slicer. In: 2004 2nd IEEE international symposium on biomedical imaging: nano to macro (IEEE Cat No. 04EX821). IEEE; 2004. p. 632–635. Çiçek Ö, et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. Springer; 2016. p. 424–432. Isola P, et al. Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017. p. 1125–1134. Van Griethuysen JJ, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Research. 2017;77(21):e104–e107. Zwanenburg A, et al. Image Biomarker Standardization Initiative. arXiv preprint arXiv:161207003. 2016. Lundberg SM, et al. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems. 2017;30. Cervantes J, Garcia-Lamont F, Rodr´ıguez-Mazahua L, Lopez A. A comprehensive survey on support vector machine classification: Applications, challenges and trends. Neurocomputing. 2020;408:189–215. Breiman L. Random forests. Machine learning. 2001;45:5–32. LeCun Y, et al. Deep learning. Nature. 2015;521(7553):436–444. Additional Declarations No competing interests reported. Supplementary Files appendix.pdf Cite Share Download PDF Status: Published Journal Publication published 11 Nov, 2025 Read the published version in BMC Medical Imaging → Version 1 posted Editorial decision: Revision requested 21 Jul, 2025 Reviews received at journal 16 Jul, 2025 Reviews received at journal 15 Jul, 2025 Reviews received at journal 14 Jul, 2025 Reviews received at journal 12 Jul, 2025 Reviewers agreed at journal 07 Jul, 2025 Reviewers agreed at journal 07 Jul, 2025 Reviewers agreed at journal 06 Jul, 2025 Reviewers agreed at journal 04 Jul, 2025 Reviewers agreed at journal 02 Jul, 2025 Reviews received at journal 25 Jun, 2025 Reviewers agreed at journal 13 Jun, 2025 Reviewers invited by journal 13 Jun, 2025 Editor assigned by journal 11 Jun, 2025 Submission checks completed at journal 11 Jun, 2025 First submitted to journal 11 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6798846","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":472014054,"identity":"01f21859-7bb4-41cb-a34f-bf592e184e5d","order_by":0,"name":"Erick Eduardo López-Ríos","email":"","orcid":"","institution":"University of Guadalajara","correspondingAuthor":false,"prefix":"","firstName":"Erick","middleName":"Eduardo","lastName":"López-Ríos","suffix":""},{"id":472014055,"identity":"4ed449e8-68c1-4f38-8503-431817a226e3","order_by":1,"name":"Francisco J. Alvarez-Padilla","email":"data:image/png;base64,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","orcid":"","institution":"University of Guadalajara","correspondingAuthor":true,"prefix":"","firstName":"Francisco","middleName":"J.","lastName":"Alvarez-Padilla","suffix":""}],"badges":[],"createdAt":"2025-06-02 05:08:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6798846/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6798846/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12880-025-01985-7","type":"published","date":"2025-11-11T15:57:34+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":84856501,"identity":"4a3ba486-5645-4ba1-a2e0-1a8188c5cd96","added_by":"auto","created_at":"2025-06-18 06:18:09","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1622763,"visible":true,"origin":"","legend":"\u003cp\u003eEffects of data augmentation on image features. (a) Comparison of the same subject's image before (blue) and after (red) augmentation. (b) Volumetric reconstruction of the thalamus, showing that random zooms can increase its apparent size even for the same image. (c) Intensity histograms from before and after augmentation, highlighting differences in distribution.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/8e896bc7265e0fd89e8060f6.png"},{"id":84856504,"identity":"07776182-8c26-494d-8ace-1e63e58164db","added_by":"auto","created_at":"2025-06-18 06:18:09","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":3122422,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentation of basal ganglia segmentation in MRI slices. Axial (a), sagittal (b), and coronal (c) views. The top row exhibits generated images, while the bottom row presents real images. Segmented structures are color-coded, including: white matter (blue), gray matter (purple), left and right caudate nuclei (green and yellow), left and right putamen (orange and pink), left and right globus pallidus (brown and cyan), and left and right thalamus (lime green and red).\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/f48719a5fd111c65841eccb6.png"},{"id":84856506,"identity":"44b4e8fd-cfad-4396-8a53-e390d3950732","added_by":"auto","created_at":"2025-06-18 06:18:09","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":340883,"visible":true,"origin":"","legend":"\u003cp\u003eNetwork architecture. The generator encoding path used a stride of 3, while the decoding path employed a stride of 2. All LeakyReLU activation layers maintained a negative slope of 0.01.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/52f7c61d0eec8a6e38d94adf.png"},{"id":84856503,"identity":"1c9deb5f-fcd9-45ff-9eac-d086e12d388d","added_by":"auto","created_at":"2025-06-18 06:18:09","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":248719,"visible":true,"origin":"","legend":"\u003cp\u003eThe proposed framework for the enhanced diagnosis of MS using generative AI and radiomics analysis.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/79e964380634314c74d0e7a2.png"},{"id":84857645,"identity":"e2c2c986-1d83-40f7-8a0b-0d20ab4e4bb3","added_by":"auto","created_at":"2025-06-18 06:26:09","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":2313196,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of real and synthetic T1-weighted MRIs. (a) Representative real samples. (b) Corresponding synthetic images generated by the cGAN. (c) Real samples transformed with data augmentation.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/59818072c14d475b3342a18c.png"},{"id":84857646,"identity":"61152150-58be-440f-855e-0328874e4f38","added_by":"auto","created_at":"2025-06-18 06:26:09","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":670646,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Violin plots of SHAP importance values computed on the original training set. (b) Violin plots of SHAP importance values computed on the combined dataset of original and GAN‑generated samples.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/ae17062e66cc7c2bbb050594.png"},{"id":96105328,"identity":"933240bb-7bbc-4d51-aaf6-e321a48edd6e","added_by":"auto","created_at":"2025-11-17 16:11:11","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":9152283,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/0b6bb495-2479-4f73-bb0e-3e8976bebe54.pdf"},{"id":84856500,"identity":"a335f3bc-5ac5-4734-87ae-685f78f709f6","added_by":"auto","created_at":"2025-06-18 06:18:09","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":60548,"visible":true,"origin":"","legend":"","description":"","filename":"appendix.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6798846/v1/cae1d5fbffd91379ccdec5a9.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"AI-Generated Data Improves Multiple Sclerosis Classification in a Basal Ganglia Radiomics Model","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eMultiple sclerosis (MS) is a chronic neurological disease that affects the central nervous system [1\u0026ndash;3]. Magnetic resonance imaging (MRI) is crucial for diagnosing MS, monitoring its progression, and assessing treatment response [4\u0026ndash;5]. Recently, subcortical structures, such as the basal ganglia, are getting recognized for their relevance in MS pathogenesis and their association with diverse clinical manifestations [6\u0026ndash;11]. Quantitative MRI analysis, leveraging radiomics, offers a powerful approach to extract high-throughput information. This information holds significant potential for developing machine learning models aimed at MS classification and prognosis [12\u0026ndash;14]. Despite its great potential, its clinical applicability and reproducibility are compromised by the quality and nature of the underlying data. Clinical datasets are inherently heterogeneous, as they are often collected from multiple centers using different scanner manufacturers, acquisition protocols, and parameter settings. This difference introduces systematic biases and noise into the images, rendering radiomic features unstable and difficult to reproduce across studies.\u003c/p\u003e \u003cp\u003eTo increase the datasets and mitigate the few samples that have been acquired with different configurations, traditional data augmentation techniques, such as rotation, flipping, scaling, and intensity adjustments have been used. However, when applied to radiomic extraction, these methods can inadvertently modify the quantitative features of the images. Even slight changes may disrupt texture statistics and histogram distributions, leading to discrepancies that fail to accurately reflect the underlying anatomical structures. These issues are exemplified in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn this context, the generation of synthetic data using GANs has emerged as a promising data augmentation strategy in medical imaging [15\u0026ndash;17]. cGANs possess the capacity to generate realistic images conditioned by specific features or classes, rendering them particularly well-suited for augmenting medical datasets. However, the application of cGANs in this domain still faces open questions concerning the fidelity of synthetic data in reflecting the full spectrum of clinical variations and the optimal methods for validating their utility in downstream analytical tasks [18].\u003c/p\u003e \u003cp\u003eIn this study, we explore the use of cGANs to synthesize T1-weighted MRI images from both MS patients and healthy controls (HC) drawn from a multi-center cohort, aiming to increase images acquired with different scanners and protocols, making the model more likely to generalize in new cohorts. Our primary objective is to assess if augmenting the training dataset with these generated images can improve the performance of a classification model based on radiomic features. These features are extracted from the segmented structures, basal ganglia, white and gray matter. We hypothesize that increasing the data variability through synthetic augmentation will mitigate the limitations imposed by small sample sizes and enhance classification accuracy. This evaluation seeks to elucidate both the strengths and limitations of this data augmentation technique.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data Description\u003c/h2\u003e \u003cp\u003eTo carry out this study, a retrospective dataset composed of T1-weighted MRI scans drawn from different mexican radiology centers was used. This dataset included a total of 166 T1-weighted MRI scans, corresponding to 144 distinct patients diagnosed with multiple sclerosis, as well as 175 scans from healthy subjects or controls. Prior to\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eanalysis, all images underwent a comprehensive preprocessing procedure. This preprocessing encompassed skull stripping, voxel spacing normalization to dimensions of 1x1x1mm, image reorientation to the RPS (Right-Posterior-Superior) coordinate system, volume cropping or padding to homogenize dimensions to 128x128x128 voxels, and finally, signal intensity normalization. This step was performed differently depending on the subsequent use of the images: for image generation training, intensities were normalized to the range \u0026minus;\u0026thinsp;1 to 1; for segmentation to the range 0 to 1, and for radiomic extraction scaled using a Z-score normalization.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Segmentation Atlas for Radiomic Extraction\u003c/h2\u003e \u003cp\u003eWe constructed a high-resolution segmentation atlas to delineate the regions of interest (ROIs) for downstream radiomic analysis, including bilateral thalamus, putamen, globus pallidus, caudate nucleus, as well as whole white and gray matter. All ROIs were initially segmented using a hybrid pipeline that combines a region‐growing algorithm with a geodesic distance‐based support vector machine (SVM) classifier [19], yielding candidate masks. Neuroanatomists subsequently performed manual refinements in 3D-Slicer [20] to ensure anatomical accuracy and consistency. Inter‐rater reliability, assessed on a subset of 20 volumes, yielded a Dice similarity coefficient of 0.92\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03 for subcortical nuclei and 0.95\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02 for white/gray matter, confirming reproducibility. An example of the resulting segmentations is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo streamline future processing, particularly for the cGAN‑generated MRI volumes, a 3D UNet [21] was then trained on these masks, enabling fully automated segmentation of synthetic images with minimal manual intervention.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Traditional Data Augmentation\u003c/h2\u003e \u003cp\u003eTo compare with data generated from GANs, we applied traditional data augmentation techniques to our input images. Images were randomly translated along a random axis with displacements ranging from \u0026minus;\u0026thinsp;10 to 10 pixels and rotated by angles between \u0026minus;\u0026thinsp;15\u0026deg; and 15\u0026deg;. Additionally, random zooming was applied to simulate variations in object scale, while Gaussian noise was introduced to mimic sensor noise. Horizontal mirroring was also performed by flipping the images along the x-axis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Image Generation\u003c/h2\u003e \u003cp\u003eTo address the limitation of data availability and investigate the potential of data augmentation, a cGAN was implemented to generate synthetic T1-weighted MRI for both MS patients and HC. The conditional aspect of the GAN was achieved by conditioning the network on the corresponding class label (MS or HC), allowing controlled generation of images specific to each group. The generator architecture, based on the UNet framework [21], and the PatchGAN-based discriminator [22] are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. After training the cGAN, 100 novel synthetic images per condition were generated, creating a synthetic dataset of comparable size and balanced classes. Finally, the generated synthetic images were used as input to a 3D UNet network for basal ganglia and white/gray matter segmentation, enabling the extraction of radiomic features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Radiomic Features\u003c/h2\u003e \u003cp\u003eThe extraction of radiomic features was performed using the PyRadiomics library [23] and followed the recommendations from the Image Biomarker Standardization Initiative (IBSI) [24]. The feature sets calculated included: Gray Level Co‑occurrence Matrix (GLCM); Gray Level Dependence Matrix (GLDM); Gray Level Run Length Matrix (GLRLM); Gray Level Size Zone Matrix (GLSZM); Shape‑based; and Statistical features.\u003c/p\u003e \u003cp\u003eFor each segmented basal ganglia structure, a total of 107 radiomic features were computed, encompassing 14 shape‑based features, 18 histogram features, 24 GLCM, 14 GLDM, 16 GLRLM, 16 GLSZM, and 5 NGTDM features (total of 1070 features per subject). To mitigate redundancy and prevent overfitting in downstream modeling, we applied a supervised feature\u0026ndash;selection pipeline based on LASSO (Least Absolute Shrinkage and Selection Operator) regression with 10‑fold cross‑validation. This process yielded a final set of 96 radiomic features. A detailed list of all the remaining features is provided in Table A2.\u003c/p\u003e \u003cp\u003eThese LASSO‑selected features were then used to train machine learning classifiers, SVM, Random Forest, and a Residual Neural Network (ResNet), on the task of distinguishing MS patients from HC. Classifier performance was evaluated in terms of accuracy, F1‑Score, sensitivity, and specificity on an independent test set.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo interpret the contribution of each radiomic descriptor to the classifier decisions, we employed SHAP (SHapley Additive exPlanations) values [25]. SHAP provides a unified measure of feature importance by attributing to each feature the change in the model\u0026rsquo;s output when conditioning on that feature\u0026rsquo;s presence. We computed SHAP values for the best‑performing model on the test cohort, generating global importance rankings of the selected radiomic features as well as individualized explanations for each subject.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Classification Algorithms\u003c/h2\u003e \u003cp\u003eIn this study, we employ a diverse set of three classification algorithms, Support Vector Machine (SVM), Random Forest (RF) and ResNet, to analyze radiomic features and to identify the most relevant predictors. For each classifier, experiments were conducted using three training configurations. In the first configuration, models were trained solely on the real dataset, while in the second, the training data were augmented with GAN images generated via cGANs. The last configuration involved the traditional data augmentation instead of the GAN-based. Note that the validation and testing sets were always composed entirely of real data to ensure an unbiased evaluation of generalization performance.\u003c/p\u003e \u003cp\u003eFor classifiers that do not require an explicit validation phase, we partitioned the real data into 70% for training and 30% for testing. However, for the ResNet a 70%/15%/15% split was used for training, validation, and testing, respectively. All models were rigorously evaluated using multiple metrics, including accuracy, F1-score, sensitivity and specificity.\u003c/p\u003e \u003cp\u003eThe SVM classifier identifies an optimal hyperplane to maximize the margin between classes and is well suited to capturing complex nonlinear patterns [26]. Its performance was optimized by tuning the kernel function, the penalty parameter \u003cem\u003eC\u003c/em\u003e, and the kernel coefficient γ via grid search with 5-fold cross-validation. The RF classifier is an ensemble method that builds and aggregates multiple decision trees, was similarly tuned by varying the number of trees and the maximum tree depth [27]. Finally, the ResNet leverages multiple residual layers of interconnected neurons to capture intricate patterns in the data [28], and its training incorporated early stopping based on validation performance. Selected hyperparameters for each classifier can be consulted in Table A1.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Implementation Details\u003c/h2\u003e \u003cp\u003eAll experiments were performed on a workstation featuring an NVIDIA L4 GPU (22.5GB VRAM, 20.7GB utilized during training) and 53GB of RAM (approximately 19.29GB in use). Python (latest version at the time) and TensorFlow (latest version) were used, along with libraries such as Nibabel, NumPy, SciPy, scikit-learn, and PyRadiomics.\u003c/p\u003e \u003cp\u003eFor the cGAN training both the generator and discriminator networks were trained using the Adam optimizer, with an initial learning rate of 0.0002 and a batch size of 4. The generator loss function was a combination of adversarial loss and L1 loss, while the discriminator loss function was binary cross-entropy loss. The output activation function for the generator network was tanh, and for the discriminator network, was sigmoid. The GAN was trained for 200 epochs. Activation functions for the generator and discriminator networks were LeakyReLU and ReLU, respectively. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates the complete pipeline of this work.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Image Generation Results\u003c/h2\u003e \u003cp\u003eThe quality of the GAN T1-weighted MRI images generated by the cGAN was quantitatively evaluated using the Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) metrics. The results revealed that the GAN images exhibit considerable structural and perceptual similarity to the real images, achieving average values of 0.9026 for SSIM and 33.82dB for PSNR. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e presents examples of the synthetic MRI images generated by the cGAN compared to real and transformed with data augmentation samples.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Classification Performance\u003c/h2\u003e \u003cp\u003eIn this study, two experiments were conducted to evaluate the performance of different classification algorithms. In Experiment 1, all radiomic features extracted from differentiated tissue regions, prior to any feature-selection procedure, were used to train the classification algorithms. In Experiment 2, only the subset of radiomic features selected via the LASSO test was used for training. Classifiers were trained using exclusively real data as well as a combination of real and traditionally augmented (TDA) or generative based (GDA) data (applied only in the training set). Tables\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e summarize the performance of the classifiers measured on the test set using the metrics of sensitivity, specificity, F1‐Score, and accuracy.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExperiment 1: Performance of classifiers using all radiomic features extracted from differentiated tissues (before LASSO).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClassifier\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8245\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6506\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9886\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8479\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8289\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.7590\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9318\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6116\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6969\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.3207\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9532\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9529\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9759\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9318\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9479\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9090\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9122\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9162\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8409\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9473\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9461\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9518\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9431\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9590\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9580\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9638\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9545\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8771\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8826\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9518\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8068\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Model Interpretability\u003c/h2\u003e \u003cp\u003eTo deepen the decision mechanisms of our classification models, we used the XAI SHAP technique to assess the global importance of radiomic features, obtained from the SHAP values calculated on the ResNet after LASSO selection, both with real and synthetic data. This analysis generates the violin plots in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, where the y-axis features reflect the contribution to the model output, allowing us to identify which ones play a more decisive role in discriminating between MS patients (positive class) and healthy controls (negative class).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eExperiment 2: Performance of classifiers using only radiomic features selected by LASSO.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClassifier\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7454\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.5882\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4761\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9117\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.7818\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.4761\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5238\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8095\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9411\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9090\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8619\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9272\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9571\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9082\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8837\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9523\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.8970\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;GDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9541\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9382\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9268\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9705\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eReal\u0026thinsp;+\u0026thinsp;TDA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9908\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9879\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9898\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.9852\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe violins illustrate the density of SHAP values, revealing in which ranges most of the observations for each feature are concentrated. Blue indicates a smaller impact on classification, while red indicates a more pronounced effect. Likewise, SHAP values greater than zero (\u0026gt;\u0026thinsp;0) correspond to features that favor classification in the positive class (EM).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eViolin plots collect SHAP values from two previously defined scenarios: (a) exclusively on the radiomic features in the real set, and (b) on the extended set with synthetic samples generated by GAN. Although each experiment provides different nuances in the significance distribution, we observe that several key features are repeated in both cases, which reinforces their predictive relevance. Prominent among these are Elongation Putamen, Surface Volume Ratio Gray Matter and Sphericity White Matter, whose SHAP distributions not only show a consistent impact on MS patient classification but also suggest that the inclusion of synthetic data does not drastically alter their influence.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eWhen synthetic images were incorporated into training, all classifiers showed gains over models trained on real data alone or with TDA. In Experiment 1 the ResNet accuracy rose from 94.61\u0026ndash;95.80% with GDA. Unlike simple geometric transforms, GAN-based augmentation appears to expose the classifier to novel radiomic heterogeneity, reducing overfitting and improving generalization.\u003c/p\u003e \u003cp\u003eIn Experiment 2, the ResNet trained on LASSO-selected features reached a F1-Score of 0.9908 with GDA, compared to 0.9382 with TDA and 0.8837 on real data alone. This underscores that adding synthetic samples not only boosts classification metrics but also sharpens the signal of the most discriminative radiomic descriptors.\u003c/p\u003e \u003cp\u003eThe SHAP analysis offered further insights into the performance. Across both real-only and GAN-augmented datasets, features such as Elongation of the Putamen, Surface-to-Volume Ratio of Gray Matter, and Sphericity of White Matter consistently ranked highest in their contributions to MS classification. Importantly, the inclusion of synthetic data did not drastically reshuffle feature importance, but rather reinforced the existing hierarchy, suggesting that GAN-generated scans preserve the key radiomic signatures of disease.\u003c/p\u003e \u003cp\u003eNevertheless, some limitations exist. First, our current work focused on cross-sectional data, whether models trained with GAN-augmented data can accurately predict future progression remains to be tested. Second, although feature stability across real and synthetic sets is promising, further testing is needed to rule out subtle biases, for instance by analyzing SHAP interactions or conducting domain adaptation studies.\u003c/p\u003e \u003cp\u003eLooking ahead, integrating multimodal GANs could further enrich virtual cohorts and enhance robustness across imaging protocols. Finally, pairing synthetic augmentation with active learning strategies could optimize when and how to request additional real scans, maximizing clinical impact while minimizing acquisition costs.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eWe have presented a flexible pipeline that combines radiomic feature extraction with GAN‑augmented MRI data to mitigate the challenges of limited and homogeneous datasets. Through comprehensive SHAP analyses, we demonstrated that a subset of texture and shape descriptors from basal ganglia structures provides consistent, interpretable insights into tissue alterations, insights that may reflect neurodegenerative changes but are not specific to multiple sclerosis alone. Rather than predicting individual disease onset, our framework highlights radiomic markers whose longitudinal shifts could guide further clinical investigation or triage for advanced imaging and biomarker testing. Crucially, these models are designed to complement, not replace clinical expertise, serving as an initial step toward integrating imaging‑based signals with diverse clinical, genetic, and biochemical data. Finally, our results confirm that GAN‑generated images can faithfully expand real‑world variability, bolster the robustness of downstream classifiers and offer a reusable resource for future radiomic and deep‑learning studies.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch3\u003eEthics approval and consent to participate\u003c/h3\u003e\n\u003cp\u003eThis study was a retrospective analysis of fully anonymized and de-identified medical images. The Hospital Civil de Guadalajara Committee determined that this research did not constitute human subjects research and waived the requirement for formal ethics committee approval and individual informed consent. The study was conducted in accordance with the principles of the Declaration of Helsinki.\u003c/p\u003e\n\u003ch3\u003eConsent for publication\u003c/h3\u003e\n\u003cp\u003eNot applicable. This study does not include any identifiable personal data or images.\u003c/p\u003e\n\u003ch3\u003eAvailability of data and materials\u003c/h3\u003e\n\u003cp\u003eThe data presented in this study are not publicly available due to internal privacy policies regarding their collection and use. However, the dataset can be made available upon reasonable request to the corresponding author, provided that the requesting party complies with the ethical and privacy considerations established by the collaborating institutions.\u003c/p\u003e\n\u003ch3\u003eCompeting interests\u003c/h3\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003ch3\u003eFunding\u003c/h3\u003e\n\u003cp\u003eThis research received no external funding.\u003c/p\u003e\n\u003ch3\u003eAuthors\u0026rsquo; contributions\u003c/h3\u003e\n\u003cp\u003eThis research was conceptualized by EE and FJ. EE led the design and implementation of the methodology, carried out the investigation, including data curation, software development, and formal analysis. EE also prepared the original draft of the manuscript and created the visualizations. FJ supervised the entire research process, guided the theoretical and methodological framing, and contributed significantly to the critical revision and refinement of the manuscript. Both authors reviewed and approved the final version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eWang L, et al. Human autoimmune diseases: a comprehensive update. Journal of Internal Medicine. 2015;278(4):369\u0026ndash;395.\u003c/li\u003e\n \u003cli\u003eKalinin I, et al. The impact of intracortical lesions on volumes of subcortical structures in multiple sclerosis. American Journal of Neuroradiology. 2020;41(5):804\u0026ndash;808.\u003c/li\u003e\n \u003cli\u003eDobson R and Giovannoni G. Multiple sclerosis\u0026ndash;a review. European Journal of Neurology. 2019;26(1):27\u0026ndash;40.\u003c/li\u003e\n \u003cli\u003eMcGinley MP, et al. Diagnosis and treatment of multiple sclerosis: a review. Jama. 2021;325(8):765\u0026ndash;779.\u003c/li\u003e\n \u003cli\u003eMurray T. Diagnosis and treatment of multiple sclerosis. BMJ. 2006;332(7540):525\u0026ndash;527.\u003c/li\u003e\n \u003cli\u003eJakimovski D, et al. Long-standing multiple sclerosis neurodegeneration: volumetric magnetic resonance imaging comparison to Parkinson\u0026rsquo;s disease, mild cognitive impairment, Alzheimer\u0026rsquo;s disease, and elderly healthy controls. Neurobiology of Aging. 2020;90:84\u0026ndash;92.\u003c/li\u003e\n \u003cli\u003eSchoonheim MM, et al. Subcortical atrophy and cognition: sex effects in multiple sclerosis. Neurology. 2012;79(17):1754\u0026ndash;1761.\u003c/li\u003e\n \u003cli\u003eAzevedo CJ, et al. Thalamic atrophy in multiple sclerosis: a magnetic resonance imaging marker of neurodegeneration throughout disease. Annals of neurology. 2018;83(2):223\u0026ndash;234.\u003c/li\u003e\n \u003cli\u003eZivadinov R, et al. Thalamic atrophy is associated with development of clinically definite multiple sclerosis. Radiology. 2013;268(3):831\u0026ndash;841.\u003c/li\u003e\n \u003cli\u003eSchoonheim MM, et al. Disability in multiple sclerosis is related to thalamic connectivity and cortical network atrophy. Multiple Sclerosis Journal. 2022;28(1):61\u0026ndash;70.\u003c/li\u003e\n \u003cli\u003eTrufanov A, et al. Basal ganglia atrophy as a marker of multiple sclerosis progression. Biomarkers in Neuropsychiatry. 2023;9:100073.\u003c/li\u003e\n \u003cli\u003eMayerhoefer ME, et al. Introduction to radiomics. Journal of Nuclear Medicine. 2020;61(4):488\u0026ndash;495.\u003c/li\u003e\n \u003cli\u003eTavakoli H, et al. Investigating the ability of radiomics features for diagnosis of the active plaque of multiple sclerosis patients. Journal of Biomedical Physics \u0026amp; Engineering. 2023;13(5):421.\u003c/li\u003e\n \u003cli\u003eLuo X, et al. Multi-lesion radiomics model for discrimination of relapsing-remitting multiple sclerosis and neuropsychiatric systemic lupus erythematosus. European Radiology. 2022;32(8):5700\u0026ndash;5710.\u003c/li\u003e\n \u003cli\u003eBarile B, et al. Data augmentation using generative adversarial neural networks on brain structural connectivity in multiple sclerosis. Computer Methods and Programs in Biomedicine. 2021;206:106113.\u003c/li\u003e\n \u003cli\u003eBrugnara G, et al. Addressing the Generalizability of AI in Radiology Using a Novel Data Augmentation Framework with Synthetic Patient Image Data: Proof-of-Concept and External Validation for Classification Tasks in Multiple Sclerosis. Radiology: artificial intelligence. 2024;6(6):e230514.\u003c/li\u003e\n \u003cli\u003eSuganthi K, et al. Review of medical image synthesis using GAN techniques. In: ITM Web of Conferences. vol. 37. EDP Sciences; 2021. p. 01005.\u003c/li\u003e\n \u003cli\u003ede Farias EC, et al. Impact of GAN-based lesion-focused medical image super-resolution on the robustness of radiomic features. Scientific Reports. 2021;11(1):21361.\u003c/li\u003e\n \u003cli\u003ePark S, et al. Seed growing for interactive image segmentation using SVM classification with geodesic distance. Electronics Letters. 2017;53(1):22\u0026ndash;24.\u003c/li\u003e\n \u003cli\u003ePieper S, et al. 3D Slicer. In: 2004 2nd IEEE international symposium on biomedical imaging: nano to macro (IEEE Cat No. 04EX821). IEEE; 2004. p. 632\u0026ndash;635.\u003c/li\u003e\n \u003cli\u003e\u0026Ccedil;i\u0026ccedil;ek \u0026Ouml;, et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention\u0026ndash;MICCAI 2016: 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II 19. Springer; 2016. p. 424\u0026ndash;432.\u003c/li\u003e\n \u003cli\u003eIsola P, et al.\u0026nbsp;Image-to-image translation with conditional adversarial networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017. p. 1125\u0026ndash;1134.\u003c/li\u003e\n \u003cli\u003eVan Griethuysen JJ, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Research. 2017;77(21):e104\u0026ndash;e107.\u003c/li\u003e\n \u003cli\u003eZwanenburg A, et al. Image Biomarker Standardization Initiative. arXiv preprint arXiv:161207003. 2016.\u003c/li\u003e\n \u003cli\u003eLundberg SM, et al. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems. 2017;30.\u003c/li\u003e\n \u003cli\u003eCervantes J, Garcia-Lamont F, Rodr\u0026acute;ıguez-Mazahua L, Lopez A. A comprehensive survey on support vector machine classification: Applications, challenges and trends. Neurocomputing. 2020;408:189\u0026ndash;215.\u003c/li\u003e\n \u003cli\u003eBreiman L. Random forests. Machine learning. 2001;45:5\u0026ndash;32.\u003c/li\u003e\n \u003cli\u003eLeCun Y, et al. Deep learning. Nature. 2015;521(7553):436\u0026ndash;444.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Generative AI, Feature Engineering, Computer-Aided Diagnosis, Multiple Sclerosis","lastPublishedDoi":"10.21203/rs.3.rs-6798846/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6798846/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e The heterogeneity and limited availability of magnetic resonance imaging (MRI) datasets in multiple sclerosis (MS) restrict the robustness of quantitative analysis methods like radiomics. This study explores the use of Generative Adversarial Networks (GANs) as a data harmonization technique. This study assesses whether GAN-generated images are realistic enough to improve the performance of radiomics-based classification models.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e We trained GANs to synthesize realistic T1w MRI from a cohort of MS patients and healthy controls. Segmented real and GAN-generated images were processed to extract statistics, texture, and shape radiomic features. Different machine-learning classifiers were trained using traditional augmentation techniques and cGAN-augmented datasets to assess pertinence. Explainable AI methods (SHAP) identified the most influential radiomic biomarkers and how they behave between real and GAN datasets.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003eGAN-generated images increased the mean classification accuracy when using a ResNet (from 0.88 to 0.98) on unseen test data. Explainability analyses revealed that texture heterogeneity and specific shape descriptors of the basal ganglia were the top predictors distinguishing MS from controls. Both datasets features show the same behavior when comparing their distributions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003eIntegrating AI-powered synthetic MRI data into radiomic pipelines substantially improves disease classification accuracy and robustness. This approach addresses data scarcity due to differences in scan protocols, uncovers imaging biomarkers, and offers a scalable strategy for enhancing clinical decision support in MS diagnostics.\u003c/p\u003e","manuscriptTitle":"AI-Generated Data Improves Multiple Sclerosis Classification in a Basal Ganglia Radiomics Model","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-18 06:18:04","doi":"10.21203/rs.3.rs-6798846/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-07-21T07:54:58+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-16T15:51:54+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-15T16:50:43+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-14T13:44:29+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-07-12T21:07:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"15757900241936473535513103586453007720","date":"2025-07-08T03:34:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"174566851995730508417255266414185526223","date":"2025-07-07T17:30:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"136923965864962774357389591045320990066","date":"2025-07-06T08:53:54+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"263025599068161950549696660916259172830","date":"2025-07-04T18:59:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"232309210707116023637890615009825973046","date":"2025-07-02T07:13:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-25T10:40:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"331635424923181907415124717642240462352","date":"2025-06-13T07:53:49+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-06-13T07:46:23+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-06-11T05:55:15+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-06-11T05:42:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-06-11T05:38:57+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"1511b2aa-f753-4c79-ba74-3d96192e72e1","owner":[],"postedDate":"June 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-11-17T16:07:12+00:00","versionOfRecord":{"articleIdentity":"rs-6798846","link":"https://doi.org/10.1186/s12880-025-01985-7","journal":{"identity":"bmc-medical-imaging","isVorOnly":false,"title":"BMC Medical Imaging"},"publishedOn":"2025-11-11 15:57:34","publishedOnDateReadable":"November 11th, 2025"},"versionCreatedAt":"2025-06-18 06:18:04","video":"","vorDoi":"10.1186/s12880-025-01985-7","vorDoiUrl":"https://doi.org/10.1186/s12880-025-01985-7","workflowStages":[]},"version":"v1","identity":"rs-6798846","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6798846","identity":"rs-6798846","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}