{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T06:17:36Z","timestamp":1787811456686,"version":"build-2784847793"},"reference-count":27,"publisher":"Springer Science and Business Media LLC","license":[{"start":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T00:00:00Z","timestamp":1787788800000},"content-version":"tdm","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0"},{"start":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T00:00:00Z","timestamp":1787788800000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0"}],"funder":[{"DOI":"10.13039\/501100003246","name":"Nederlandse Organisatie voor Wetenschappelijk Onderzoek","doi-asserted-by":"publisher","award":["NGF.1607.22.032"],"award-info":[{"award-number":["NGF.1607.22.032"]}],"id":[{"id":"10.13039\/501100003246","id-type":"DOI","asserted-by":"publisher"}]},{"DOI":"10.13039\/100018696","name":"HORIZON EUROPE Health","doi-asserted-by":"publisher","award":["101079894"],"award-info":[{"award-number":["101079894"]}],"id":[{"id":"10.13039\/100018696","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["link.springer.com"],"crossmark-restriction":false},"short-container-title":["Int J CARS"],"abstract":"<jats:title>Abstract<\/jats:title>\n                  <jats:sec>\n                    <jats:title>Purpose<\/jats:title>\n                    <jats:p>Foundation models are increasingly used as frozen feature extractors for CT classification tasks, yet the determinants of their downstream performance remain unclear. We assess how foundation model choice, feature aggregation strategy, and abnormality type are associated with performance in organ-level abnormality classification. We further evaluate whether more expressive aggregation strategies outperform simple pooling, and whether performance varies by abnormality type.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>We evaluate five state-of-the-art medical imaging foundation models on binary organ-level abnormality classification (normal vs. abnormal) across six abdominal organs. Local patch-level embeddings are aggregated using six strategies, including simple pooling methods (e.g., mean pooling) and attention-based multiple instance learning. A linear classifier is trained on feature embeddings, and performance is evaluated on an annotated test set of 200 CT scans.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>\n                      Among the evaluated models, 3D CT-native models generally outperformed 2D multi-modal models. None of the tested aggregation strategies significantly outperformed mean pooling, including attention-based multiple instance learning (best-performing aggregation vs. mean:\n                      <jats:inline-formula>\n                        <jats:alternatives>\n                          <jats:tex-math>$$\\Delta $$<\/jats:tex-math>\n                          <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\">\n                            <mml:mi>\u0394<\/mml:mi>\n                          <\/mml:math>\n                        <\/jats:alternatives>\n                      <\/jats:inline-formula>\n                      AUC = 0.008, 95% CI\n                      <jats:inline-formula>\n                        <jats:alternatives>\n                          <jats:tex-math>$$[-0.002, 0.022]$$<\/jats:tex-math>\n                          <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\">\n                            <mml:mrow>\n                              <mml:mo>[<\/mml:mo>\n                              <mml:mo>-<\/mml:mo>\n                              <mml:mn>0.002<\/mml:mn>\n                              <mml:mo>,<\/mml:mo>\n                              <mml:mn>0.022<\/mml:mn>\n                              <mml:mo>]<\/mml:mo>\n                            <\/mml:mrow>\n                          <\/mml:math>\n                        <\/jats:alternatives>\n                      <\/jats:inline-formula>\n                      ). In our organ-level classification setting, performance differed between abnormality type for all well-performing models (SPECTRE, TAP-CT, CT-FM), with lower AUCs observed for focal abnormalities compared to diffuse abnormalities (largest difference:\n                      <jats:inline-formula>\n                        <jats:alternatives>\n                          <jats:tex-math>$$\\Delta $$<\/jats:tex-math>\n                          <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\">\n                            <mml:mi>\u0394<\/mml:mi>\n                          <\/mml:math>\n                        <\/jats:alternatives>\n                      <\/jats:inline-formula>\n                      AUC = 0.108, 95% CI [0.080,\u00a00.135]). This gap was not reduced by the evaluated aggregation strategies.\n                    <\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Conclusion<\/jats:title>\n                    <jats:p>In our experiments, downstream performance varied more across foundation models than across aggregation strategies. The choice of aggregation strategy, including attention-based multiple instance learning, did not significantly impact performance in this setting. The persistent gap for focal abnormalities suggests that current representations may insufficiently encode localized disease patterns, motivating the development of localization-aware pre-training approaches.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.1007\/s11548-026-03786-x","type":"journal-article","created":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T05:25:28Z","timestamp":1787808328000},"update-policy":"https:\/\/doi.org\/10.1007\/springer_crossmark_policy","source":"Crossref","is-referenced-by-count":0,"title":["Systematic evaluation of foundation models for organ-level classification on CT scans"],"prefix":"10.1007","author":[{"ORCID":"https:\/\/orcid.org\/0009-0007-0047-9225","authenticated-orcid":false,"given":"Jan","family":"Tagscherer","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-5184-4340","authenticated-orcid":false,"given":"Sarah","family":"de Boer","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-9781-4101","authenticated-orcid":false,"given":"Fennie","family":"van der Graaf","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0005-2742-7041","authenticated-orcid":false,"given":"Lena","family":"Philipp","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0003-1180-3805","authenticated-orcid":false,"given":"Colin","family":"Jacobs","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-7090-2742","authenticated-orcid":false,"given":"Ewoud J.","family":"Smit","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-7602-803X","authenticated-orcid":false,"given":"Alessa","family":"Hering","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"297","published-online":{"date-parts":[[2026,8,27]]},"reference":[{"key":"3786_CR1","doi-asserted-by":"publisher","unstructured":"Bai H, Ge C, Li JY, Luo Z, Ma P, Wan W, Yang X, Zhang J, Zhanng R, Zhu L, Y (2022) Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. In: Advances in neural information processing systems 35. Neural information processing systems foundation, Inc. (NeurIPS). NeurIPS, pp 36722\u201336732. https:\/\/doi.org\/10.52202\/068431-2661","DOI":"10.52202\/068431-2661"},{"key":"3786_CR2","doi-asserted-by":"crossref","unstructured":"Bigolin\u00a0Lanfredi R, Zhuang Y, Finkelstein M, Thoppey Srinivasan\u00a0Balamuralikrishna P, Krembs L, Khoury B, Reddy A, Mukherjee P, Rofsky NM, Summers RM (2025) Leavs: an llm-based labeler for abdominal ct supervision. In: International conference on medical image computing and computer-assisted intervention. Springer, pp 316\u2013326","DOI":"10.1007\/978-3-032-04971-1_30"},{"issue":"8112","key":"3786_CR3","doi-asserted-by":"publisher","first-page":"1318","DOI":"10.1038\/s41586-026-10181-8","volume":"652","author":"L Blankemeier","year":"2026","unstructured":"Blankemeier L, Kumar A, Cohen JP et al (2026) Merlin: a computed tomography vision-language foundation model and dataset. Nature 652(8112):1318\u20131328. https:\/\/doi.org\/10.1038\/s41586-026-10181-8","journal-title":"Nature"},{"key":"3786_CR4","doi-asserted-by":"publisher","DOI":"10.1016\/j.compbiomed.2025.111321","volume":"199","author":"JS Bosma","year":"2025","unstructured":"Bosma JS, Builtjes L, Saha A, Twilt JJ, Tsiknakis M, Marias K, Regge D, Papanikolaou N, Schoots IG, Veltman J, Elschot M, Yakar D, Obuchowski NA, Heinrich MP, Hering A, de Rooij M, Huisman H (2025) Scalable clinical annotation with location evidence (scale). Comput Biol Med 199:111321. https:\/\/doi.org\/10.1016\/j.compbiomed.2025.111321","journal-title":"Comput Biol Med"},{"issue":"1","key":"3786_CR5","doi-asserted-by":"publisher","first-page":"172","DOI":"10.1148\/radiol.220152","volume":"306","author":"PT Chen","year":"2023","unstructured":"Chen PT, Wu T, Wang P, Chang D, Liu KL, Wu MS, Roth HR, Lee PC, Liao WC, Wang W (2023) Pancreatic cancer detection on ct scans with deep learning: a nationwide population-based study. Radiology 306(1):172\u2013182. https:\/\/doi.org\/10.1148\/radiol.220152","journal-title":"Radiology"},{"key":"3786_CR6","unstructured":"Chen T, Kornblith S, Norouzi M, Hinton G (2020) A simple framework for contrastive learning of visual representations. In: International conference on machine learning. pp 1597\u20131607 (PMLR)"},{"key":"3786_CR7","unstructured":"Claessens C, Viviers C, D\u2019Amicantonio G, Bondarev E, van\u00a0der Sommen F (2026) Scaling self-supervised and cross-modal pretraining for volumetric ct transformers. In: Proceedings of the IEEE\/CVF conference on computer vision and pattern recognition, pp 13636\u201313647"},{"key":"3786_CR8","unstructured":"Dancette C, Khlaut J, Saporta A, Philippe H, Ferreres E, Callard B, Danielou T, Alberge L, Machado L, Tordjman D, Dupuis J, Floch L, K, Terrail D, Moshiri J, Dercle M, Boeken L, Gregory T, Ronot J, Legou M, Roux F, Sapoval P, Manceron M, H\u00e9rent P, P (2025) Curia: a multi-modal foundation model for radiology. arXiv preprint. arXiv:2509.06830"},{"key":"3786_CR9","doi-asserted-by":"publisher","first-page":"129889","DOI":"10.1109\/access.2020.3006362","volume":"8","author":"X Dong","year":"2020","unstructured":"Dong X, Zhou Y, Wang L, Peng J, Lou Y, Fan Y (2020) Liver cancer detection using hybridized fully convolutional neural network based on deep learning framework. IEEE Access 8:129889\u2013129898. https:\/\/doi.org\/10.1109\/access.2020.3006362","journal-title":"IEEE Access"},{"key":"3786_CR10","doi-asserted-by":"publisher","DOI":"10.1016\/j.media.2020.101857","volume":"67","author":"RL Draelos","year":"2021","unstructured":"Draelos RL, Dov D, Mazurowski MA, Lo JY, Henao R, Rubin GD, Carin L (2021) Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Med Image Anal 67:101857. https:\/\/doi.org\/10.1016\/j.media.2020.101857","journal-title":"Med Image Anal"},{"key":"3786_CR11","unstructured":"Hendrycks D, Burns C, Basart S, Zou A, Mazeika M, Song D, Steinhardt J (2021) Measuring massive multitask language understanding. In: Proceedings of the international conference on learning representations (ICLR)"},{"key":"3786_CR12","unstructured":"Ilse M, Tomczak J, Welling M (2018) Attention-based deep multiple instance learning. In: International conference on machine learning. pp 2127\u20132136 (PMLR)"},{"key":"3786_CR13","doi-asserted-by":"crossref","unstructured":"Lei W, Tian H, Dai L, Chen H, Zhang X (2025) Lesiondiffusion: towards text-controlled general lesion synthesis. In: International conference on medical image computing and computer-assisted intervention. Springer, pp 327\u2013336","DOI":"10.1007\/978-3-032-04971-1_31"},{"key":"3786_CR14","doi-asserted-by":"publisher","DOI":"10.1007\/s10278-025-01747-5","author":"X Li","year":"2025","unstructured":"Li X, Merkow J, Codella NCF, Santamaria-Pang A, Sangani N, Ersoy A, Burt C, Garrett JW, Bruce RJ, Warner JD, Bradshaw T, Tarapov I, Lungren MP, McMillan AB (2025) From embeddings to accuracy: comparing foundation models for radiographic classification. J Imaging Inf Med. https:\/\/doi.org\/10.1007\/s10278-025-01747-5","journal-title":"J Imaging Inf Med"},{"key":"3786_CR15","doi-asserted-by":"publisher","unstructured":"Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F (2021) Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng 5(6):555\u2013570. https:\/\/doi.org\/10.1038\/s41551-020-00682-w","DOI":"10.1038\/s41551-020-00682-w"},{"key":"3786_CR16","unstructured":"Oquab M, Darcet T, Moutakanni T, Vo HV, Szafraniec M, Khalidov V, Fernandez P, Haziza D, Massa F, El-Nouby A, Assran M, Ballas N, Galuba W, Howes R, Huang PY, Li SW, Misra I, Rabbat M, Sharma V, Synnaeve G, Xu H, Jegou H, Mairal J, Labatut P, Joulin A, Bojanowski P (2024) DINOv2: learning robust visual features without supervision. Trans Mach Learn Res"},{"issue":"3","key":"3786_CR17","doi-asserted-by":"publisher","first-page":"354","DOI":"10.1038\/s42256-024-00807-9","volume":"6","author":"S Pai","year":"2024","unstructured":"Pai S, Bontempi D, Hadzic I, Prudente V, Soka\u010d M, Chaunzwa TL, Bernatz S, Hosny A, Mak RH, Birkbak NJ, Aerts HJWL (2024) Foundation model for cancer imaging biomarkers. Nat Mach Intell 6(3):354\u2013367","journal-title":"Nat Mach Intell"},{"key":"3786_CR18","unstructured":"Pai S, Hadzic I, Bontempi D, Bressem K, Kann BH, Fedorov A, Mak RH, Aerts HJ (2025) Vision foundation models for computed tomography. arXiv preprint arXiv:2501.09001"},{"key":"3786_CR19","doi-asserted-by":"publisher","unstructured":"Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, Berg AC, Fei-Fei L (2015) Imagenet large scale visual recognition challenge. Int J Comput Vis 115(3):211\u2013252. https:\/\/doi.org\/10.1007\/s11263-015-0816-y","DOI":"10.1007\/s11263-015-0816-y"},{"issue":"7","key":"3786_CR20","doi-asserted-by":"publisher","first-page":"495","DOI":"10.1038\/s43588-024-00662-z","volume":"4","author":"R Sch\u00e4fer","year":"2024","unstructured":"Sch\u00e4fer R, Nicke T, H\u00f6fener H, Lange A, Merhof D, Feuerhake F, Schulz V, Lotz J, Kiessling F (2024) Overcoming data scarcity in biomedical imaging with a foundational multi-task model. Nat Comput Sci 4(7):495\u2013509. https:\/\/doi.org\/10.1038\/s43588-024-00662-z","journal-title":"Nat Comput Sci"},{"key":"3786_CR21","unstructured":"Stegeman M, Philipp L, van\u00a0der Graaf F, D\u2019Amato M, Grisi C, Builtjes L, Bosma JS, Lefkes J, Weber RA, Meakin JA, Koopman T, Mickan A, Prokop M, Smit EJ, Litjens G, van\u00a0der Laak J, van Ginneken B, de\u00a0Rooij M, Huisman H, Jacobs C, Ciompi F, Hering A (2026) Designing unicorn: a unified benchmark for imaging in computational pathology, radiology, and natural language. arXiv preprint arXiv:2603.02790"},{"key":"3786_CR22","doi-asserted-by":"crossref","unstructured":"Tagscherer J, de\u00a0Boer S, Philipp L, van\u00a0der Graaf F, Peeters D, Bosma J, Leijten L, Obreja B, Smit E, Hering A (2026) Evalblocks: A modular pipeline for rapidly evaluating foundation models in medical imaging. In: BVM workshop, Springer, pp 48\u201354","DOI":"10.1007\/978-3-658-51100-5_9"},{"key":"3786_CR23","unstructured":"Veenboer T, Yiasemis G, Marcus E, Van\u00a0Veldhuizen V, Snoek CG, Teuwen J, Lipman KB (2026) TAP-CT: 3d task-agnostic pretraining of computed tomography foundation models. In: Medical imaging with deep learning"},{"key":"3786_CR24","doi-asserted-by":"publisher","DOI":"10.1038\/s41597-023-02460-0","author":"D Wang","year":"2023","unstructured":"Wang D, Wang X, Wang L, Li M, Da Q, Liu X, Gao X, Shen J, He J, Shen T, Duan Q, Zhao J, Li K, Qiao Y, Zhang S (2023) A real-world dataset and benchmark for foundation model adaptation in medical image classification. Sci Data. https:\/\/doi.org\/10.1038\/s41597-023-02460-0","journal-title":"Sci Data"},{"key":"3786_CR25","doi-asserted-by":"publisher","DOI":"10.1148\/ryai.230024","author":"J Wasserthal","year":"2023","unstructured":"Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, Bach M, Segeroth M (2023) Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Artif Intell Radiol. https:\/\/doi.org\/10.1148\/ryai.230024","journal-title":"Artif Intell Radiol"},{"issue":"03","key":"3786_CR26","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1117\/1.jmi.5.3.036501","volume":"5","author":"K Yan","year":"2018","unstructured":"Yan K, Wang X, Lu L, Summers RM (2018) Deeplesion: automated mining of large-scale lesion annotations and universal lesion detection with deep learning. J Med Imaging 5(03):1. https:\/\/doi.org\/10.1117\/1.jmi.5.3.036501","journal-title":"J Med Imaging"},{"key":"3786_CR27","doi-asserted-by":"publisher","DOI":"10.1016\/j.media.2023.102996","volume":"91","author":"S Zhang","year":"2024","unstructured":"Zhang S, Metaxas D (2024) On the challenges and perspectives of foundation models for medical image analysis. Med Image Anal 91:102996. https:\/\/doi.org\/10.1016\/j.media.2023.102996","journal-title":"Med Image Anal"}],"container-title":["International Journal of Computer Assisted Radiology and Surgery"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/link.springer.com\/content\/pdf\/10.1007\/s11548-026-03786-x.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/link.springer.com\/article\/10.1007\/s11548-026-03786-x","content-type":"text\/html","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/link.springer.com\/content\/pdf\/10.1007\/s11548-026-03786-x.pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,8,27]],"date-time":"2026-08-27T05:25:32Z","timestamp":1787808332000},"score":1,"resource":{"primary":{"URL":"https:\/\/link.springer.com\/10.1007\/s11548-026-03786-x"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,27]]},"references-count":27,"alternative-id":["3786"],"URL":"https:\/\/doi.org\/10.1007\/s11548-026-03786-x","relation":{},"ISSN":["1861-6429"],"issn-type":[{"value":"1861-6429","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,27]]},"assertion":[{"value":"7 April 2026","order":1,"name":"received","label":"Received","group":{"name":"ArticleHistory","label":"Article History"}},{"value":"14 August 2026","order":2,"name":"accepted","label":"Accepted","group":{"name":"ArticleHistory","label":"Article History"}},{"value":"27 August 2026","order":3,"name":"first_online","label":"First Online","group":{"name":"ArticleHistory","label":"Article History"}},{"value":"C.J. reports receiving research grants and royalties to his host institution from MeVis Medical Solutions (Bremen, Germany), and lecture fees from Canon Medical Systems and Johnson & Johnson. He is also involved in a public-private research collaboration between Radboudumc and Philips Medical Systems (Best, the Netherlands). F.v.d.G. reports involvement in a public-private project funded in part by MeVis Medical Solutions. L.P. reports that her PhD is funded by Siemens Healthineers. The remaining authors declare that they have no conflict of interest.","order":1,"name":"Ethics","label":"Conflict of interest","group":{"name":"EthicsHeading","label":"Declarations"}},{"value":"This study used only publicly available, anonymized datasets (AMOS-MM with LEAVS annotations). No ethics approval was required.","order":2,"name":"Ethics","label":"Ethics approval","group":{"name":"EthicsHeading","label":"Declarations"}},{"value":"This study used only publicly available, anonymized data. No human participants were directly involved, and informed consent was not required.","order":3,"name":"Ethics","label":"Consent to participate","group":{"name":"EthicsHeading","label":"Declarations"}},{"value":"The manuscript does not contain any identifiable individual data. Consent for publication is therefore not applicable.","order":4,"name":"Ethics","label":"Consent to publish","group":{"name":"EthicsHeading","label":"Declarations"}}]}}