Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data

Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data | 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 Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data Stefania L. Moroianu, Christian Bluethgen, Pierre Chambon, Mehdi Cherti, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7687810/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Achieving robust performance and fairness across diverse patient populations remains a central challenge in developing clinically deployable deep learning models for diagnostic imaging. Synthetic data generation has emerged as a promising strategy to address current limitations in dataset scale and diversity. In this study, we introduce RoentGen-v2 , a state-of-the-art text-to-image diffusion model for chest radiographs that enables fine-grained control over both radiographic findings and patient demographic attributes, including sex, age, and race/ethnicity. RoentGen-v2 is the first model to generate clinically plausible chest radiographs with explicit demographic conditioning, facilitating the creation of a large, demographically balanced synthetic dataset comprising over 565,000 images. We use this large synthetic dataset to evaluate optimal training pipelines for downstream disease classification models. In contrast to prior work that combines real and synthetic data naively, we propose an improved training strategy that leverages synthetic data for supervised pretraining, followed by fine-tuning on real data. Through extensive evaluation on over 137,000 held-out chest radiographs from five institutions, we demonstrate that synthetic pretraining consistently improves model performance, generalization to out-of-distribution settings, and fairness across demographic subgroups defined across varying fairness metrics. Across datasets, synthetic pretraining led to a 6.5% accuracy increase in the performance of downstream classification models, compared to a modest 2.7% increase when naively combining real and synthetic data. We observe this performance improvement simultaneously with the reduction of the underdiagnosis fairness gap by 19.3%, with marked improvements across intersectional subgroups of sex, age, and race/ethnicity. Our proposed data-centric training approach that combines high-fidelity synthetic training data with multi-stage training pipelines is label-efficient, reducing reliance on large quantities of annotated real data. These results highlight the potential of demographically controllable synthetic imaging to advance equitable and generalizable medical deep learning under real-world data constraints. We open source our code, trained models, and synthetic dataset. Full Text Additional Declarations Competing interest reported. C.B. received research support from Promedica Foundation, Chur, Switzerland. P.C. is a researcher at Meta; his role is unrelated to the content of this study. J.G. receives research support from NIH grants R01HL167811, 1R25OD039834-01. J.G. is a member of the following: Advisory Board – AHA debiasing clinical care algorithms (DECCA); Council of medical specialty societies Encoding Equity Initiative; American College of Radiology AI Advisory Council; Board member - Society of Imaging Informatics in Medicine (SIIM); Associate editor – RSNA AI Journal Trainee Editorial Board. Unrelated to this work, J.G. received a speaker fee from Cook Medical. C.P.L. reports activities not related to the present article: Board of directors and shareholder, Bunkerhill Health. Option holder, whiterabbit.ai. Advisor and option holder, GalileoCDS. Advisor and option holder, Sirona Medical. Advisor and option holder, Adra. Advisor and option holder, Cognita. Advisor and option holder, TurboRadiology. Paid consultant, Sixth Street. Speaker fee, McKinsey and Company. Speaker fee, Philips. Recent grant and gift support paid to C.P.L.'s institution: Amazon Web Services, BunkerHill Health, Carestream, CARPL, Clairity, GE Healthcare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai. Unrelated to this work, A.S.C. receives research support from GE Healthcare, Philips, Microsoft, Amazon, Google, NVIDIA, Stability; has provided consulting services to Patient Square Capital, Chondrometrics GmbH, and Elucid Bioimaging; is co-founder of Cognita; has equity interest in Cognita, Subtle Medical, LVIS Corp, Brain Key. The other authors declare no competing interests. Supplementary Files RoentGenv2Supplementary.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 20 Nov, 2025 Reviewers agreed at journal 18 Nov, 2025 Reviewers agreed at journal 15 Nov, 2025 Reviewers invited by journal 13 Nov, 2025 Editor assigned by journal 11 Nov, 2025 Editor invited by journal 16 Oct, 2025 Submission checks completed at journal 15 Oct, 2025 First submitted to journal 15 Oct, 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-7687810","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":549379335,"identity":"2ab70fe5-2ef8-451f-8fbb-468934247cd4","order_by":0,"name":"Stefania L. Moroianu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYDACdjBpAyY/MDAwA6kEAlqYwWQaAw8DA+MMUrQcJkGLfDN34uOKivP29uxnDzZ83GPNwM+eY4BXi8Fh3s2GZ87cTuzhyUtsnPEsnUGy5w0BLcy82yQb224n8DDkmD/mOXCYweAGAVvkm8Faztnz8L8xbAZpsSekheEwWMsBxh6JHIgWAwli/NJwJjmx58Ybw8YZB9J5JM48K8DvsPbejQ8bKuzs2ftzDBs+HLCW429P3oDfYeiAhzTlo2AUjIJRMAqwAgAzwkTiIcS9KAAAAABJRU5ErkJggg==","orcid":"","institution":"Stanford University","correspondingAuthor":true,"prefix":"","firstName":"Stefania","middleName":"L.","lastName":"Moroianu","suffix":""},{"id":549379336,"identity":"c7f21188-a5cd-4360-a5d1-73fb1a4bbf13","order_by":1,"name":"Christian Bluethgen","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Christian","middleName":"","lastName":"Bluethgen","suffix":""},{"id":549379337,"identity":"6417759e-7c93-46f4-b21f-26e75f96bf2d","order_by":2,"name":"Pierre Chambon","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Pierre","middleName":"","lastName":"Chambon","suffix":""},{"id":549379338,"identity":"141ca792-20d3-43c0-9a35-90043ea662aa","order_by":3,"name":"Mehdi Cherti","email":"","orcid":"","institution":"Forschungszentrum Jülich","correspondingAuthor":false,"prefix":"","firstName":"Mehdi","middleName":"","lastName":"Cherti","suffix":""},{"id":549379339,"identity":"e79abb9e-5e72-4085-afd5-1331311e33f5","order_by":4,"name":"Jean-Benoit Delbrouck","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Jean-Benoit","middleName":"","lastName":"Delbrouck","suffix":""},{"id":549379340,"identity":"7dc271fc-507d-4f2a-9c10-2f7762a0def8","order_by":5,"name":"Magdalini Paschali","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Magdalini","middleName":"","lastName":"Paschali","suffix":""},{"id":549379341,"identity":"cf5eab5c-4f56-4df6-a9a5-a507768e7d54","order_by":6,"name":"Brandon Price","email":"","orcid":"","institution":"Emory University","correspondingAuthor":false,"prefix":"","firstName":"Brandon","middleName":"","lastName":"Price","suffix":""},{"id":549379342,"identity":"0690e11b-bf57-41b3-9def-2baee3155df8","order_by":7,"name":"Judy Gichoya","email":"","orcid":"","institution":"Emory University","correspondingAuthor":false,"prefix":"","firstName":"Judy","middleName":"","lastName":"Gichoya","suffix":""},{"id":549379343,"identity":"84c356c0-54ca-4713-99cd-c0105a4af402","order_by":8,"name":"Jenia Jitsev","email":"","orcid":"","institution":"Forschungszentrum Jülich","correspondingAuthor":false,"prefix":"","firstName":"Jenia","middleName":"","lastName":"Jitsev","suffix":""},{"id":549379344,"identity":"886e7bac-f3c3-4ec6-80db-a86fe618d13c","order_by":9,"name":"Curtis P. Langlotz","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Curtis","middleName":"P.","lastName":"Langlotz","suffix":""},{"id":549379345,"identity":"32fff3e8-8197-481b-bd9e-0b2f122aa2ea","order_by":10,"name":"Akshay S. Chaudhari","email":"","orcid":"","institution":"Stanford University","correspondingAuthor":false,"prefix":"","firstName":"Akshay","middleName":"S.","lastName":"Chaudhari","suffix":""}],"badges":[],"createdAt":"2025-09-23 02:08:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7687810/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7687810/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":96589108,"identity":"6d64929a-1a15-4bef-bec7-086110c55b9c","added_by":"auto","created_at":"2025-11-24 05:55:33","extension":"json","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16361,"visible":true,"origin":"","legend":"","description":"","filename":"be393985771e42a2b5328c779f2317a3.json","url":"https://assets-eu.researchsquare.com/files/rs-7687810/v1/090e93f673b856ae3864f053.json"},{"id":96605446,"identity":"2dac58cd-f143-4080-bfb9-c447cf9d6301","added_by":"auto","created_at":"2025-11-24 09:23:04","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":13895069,"visible":true,"origin":"","legend":"","description":"","filename":"RoentGenV2ManuscriptBMC1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7687810/v1_covered_2dca126a-3ba4-41af-b5f1-e148903f9c9a.pdf"},{"id":96589109,"identity":"68f4f367-732c-48aa-91d3-9c06ee6b2658","added_by":"auto","created_at":"2025-11-24 05:55:33","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":6267186,"visible":true,"origin":"","legend":"","description":"","filename":"RoentGenv2Supplementary.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7687810/v1/b34cd7cde901d660a4966001.pdf"}],"financialInterests":"Competing interest reported. C.B. received research support from Promedica Foundation, Chur, Switzerland.\nP.C. is a researcher at Meta; his role is unrelated to the content of this study.\nJ.G. receives research support from NIH grants R01HL167811, 1R25OD039834-01.\nJ.G. is a member of the following: Advisory Board – AHA debiasing clinical care algorithms (DECCA); Council of medical specialty societies Encoding Equity Initiative; American College of Radiology AI Advisory Council; Board member - Society of Imaging Informatics in Medicine (SIIM); Associate editor – RSNA AI Journal Trainee Editorial Board. Unrelated to this work, J.G. received a speaker fee from Cook Medical.\nC.P.L. reports activities not related to the present article: Board of directors and shareholder, Bunkerhill Health. Option holder, whiterabbit.ai. Advisor and option holder, GalileoCDS. Advisor and option holder, Sirona Medical. Advisor and option holder, Adra. Advisor and option holder, Cognita. Advisor and option holder, TurboRadiology. Paid consultant, Sixth Street. Speaker fee, McKinsey and Company. Speaker fee, Philips.\nRecent grant and gift support paid to C.P.L.'s institution: Amazon Web Services, BunkerHill Health, Carestream, CARPL, Clairity, GE Healthcare, Google Cloud, IBM, Kheiron, Lambda, Lunit, Microsoft, Nightingale Open Science, Philips, Siemens Healthineers, Stability.ai, Subtle Medical, VinBrain, Visiana, Whiterabbit.ai.\nUnrelated to this work, A.S.C. receives research support from GE Healthcare, Philips, Microsoft, Amazon, Google, NVIDIA, Stability; has provided consulting services to Patient Square Capital, Chondrometrics GmbH, and Elucid Bioimaging; is co-founder of Cognita; has equity interest in Cognita, Subtle Medical, LVIS Corp, Brain Key.\n\nThe other authors declare no competing interests.","formattedTitle":"Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"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":"","lastPublishedDoi":"10.21203/rs.3.rs-7687810/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7687810/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Achieving robust performance and fairness across diverse patient populations remains a central challenge in developing clinically deployable deep learning models for diagnostic imaging. Synthetic data generation has emerged as a promising strategy to address current limitations in dataset scale and diversity. In this study, we introduce RoentGen-v2 , a state-of-the-art text-to-image diffusion model for chest radiographs that enables fine-grained control over both radiographic findings and patient demographic attributes, including sex, age, and race/ethnicity. RoentGen-v2 is the first model to generate clinically plausible chest radiographs with explicit demographic conditioning, facilitating the creation of a large, demographically balanced synthetic dataset comprising over 565,000 images. We use this large synthetic dataset to evaluate optimal training pipelines for downstream disease classification models. In contrast to prior work that combines real and synthetic data naively, we propose an improved training strategy that leverages synthetic data for supervised pretraining, followed by fine-tuning on real data. Through extensive evaluation on over 137,000 held-out chest radiographs from five institutions, we demonstrate that synthetic pretraining consistently improves model performance, generalization to out-of-distribution settings, and fairness across demographic subgroups defined across varying fairness metrics. Across datasets, synthetic pretraining led to a 6.5% accuracy increase in the performance of downstream classification models, compared to a modest 2.7% increase when naively combining real and synthetic data. We observe this performance improvement simultaneously with the reduction of the underdiagnosis fairness gap by 19.3%, with marked improvements across intersectional subgroups of sex, age, and race/ethnicity. Our proposed data-centric training approach that combines high-fidelity synthetic training data with multi-stage training pipelines is label-efficient, reducing reliance on large quantities of annotated real data. These results highlight the potential of demographically controllable synthetic imaging to advance equitable and generalizable medical deep learning under real-world data constraints. We open source our code, trained models, and synthetic dataset.","manuscriptTitle":"Improving Performance, Robustness, and Fairness of Radiographic AI Models with Finely-Controllable Synthetic Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-24 05:55:29","doi":"10.21203/rs.3.rs-7687810/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2025-11-20T16:25:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"331635424923181907415124717642240462352","date":"2025-11-18T13:01:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"333527940252009697450471634625509316450","date":"2025-11-15T12:01:02+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-13T11:52:55+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-11T10:05:17+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-16T06:15:58+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-15T20:00:30+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2025-10-15T19:57:55+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":"c120b133-c8c3-464a-b2dd-119cd7df80da","owner":[],"postedDate":"November 24th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-27T22:23:21+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-24 05:55:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7687810","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7687810","identity":"rs-7687810","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}