pith. sign in

arxiv: 2606.05460 · v1 · pith:L7PNC674new · submitted 2026-06-03 · 💻 cs.CV

ORACLE-CT: Anatomy-Aware Support Pooling for CT Classification

Lavsen Dahal , Yubraj Bhandari , Geoffrey Rubin , Joseph Y. Lo This is my paper

Pith reviewed 2026-06-28 06:23 UTC · model grok-4.3

classification 💻 cs.CV
keywords CT classificationanatomy-aware poolingmulti-organ segmentationsupport-masked poolingabdominal CTencoder aggregationdisease classificationexternal validation
0
0 comments X

The pith

Anatomy-aware support pooling using multi-organ segmentation improves abdominal CT disease classification over global average pooling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces ORACLE-CT, a framework that defines label-specific anatomical supports from multi-organ segmentation and restricts attention pooling to those regions instead of pooling across the entire volume. This addresses the mismatch between localized disease evidence in specific organs and the global aggregation used in most CT classifiers. Experiments across DINOv3, I3D-ResNet-121, and Pillar-0 encoders show consistent gains in macro-AUROC and AUPRC on the MERLIN training set and improved transfer performance on external Duke-Abdomen and AMOS evaluations. The method supports multiple pooling strategies while keeping predictions linked to explicit anatomical regions.

Core claim

ORACLE-CT is an encoder-agnostic aggregation framework that uses multi-organ segmentation to create anatomical supports and applies support-masked pooling (single-organ, union, comparative, localized, or global) to restrict feature aggregation to relevant compartments, yielding higher macro-AUROC/AUPRC than global average pooling on MERLIN (DINOv3: 0.858/0.676 vs 0.838/0.638) and better external results on harmonized 10-label sets.

What carries the argument

Support-masked pooling, which uses multi-organ segmentation to define label-specific anatomical supports and restricts attention pooling to those regions.

If this is right

  • Support-masked pooling raises performance for DINOv3 and I3D-ResNet-121 on both internal MERLIN metrics and external Duke-Abdomen/AMOS transfer.
  • For the Pillar-0 encoder, most improvement comes from learned attention with smaller added benefit from anatomical masking.
  • The framework preserves an auditable link between each prediction and the specific anatomical supports used.
  • Multiple support strategies (single-organ, multi-organ union, comparative) are compatible with the same encoder backbone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The gains suggest that many CT findings are sufficiently localized that global pooling dilutes the signal even when attention is learned.
  • If segmentation quality varies across sites, the method may require domain-specific fine-tuning of the segmenter before pooling is applied.
  • The approach could be extended to other 3D modalities provided reliable organ masks exist, though the paper does not test this.

Load-bearing premise

Accurate multi-organ segmentation is available and the segmented regions contain all diagnostic evidence without missing important information elsewhere in the scan.

What would settle it

On a dataset where disease evidence lies outside the segmented organs, support-masked pooling produces equal or lower AUROC than global average pooling.

Figures

Figures reproduced from arXiv: 2606.05460 by Geoffrey Rubin, Joseph Y. Lo, Lavsen Dahal, Yubraj Bhandari.

Figure 1
Figure 1. Figure 1: Conceptual overview of ORACLE–CT. (A) Standard study-level classification commonly pools over the full spatial feature lattice, potentially mixing pathology-relevant and irrelevant anatomy. (B) ORACLE–CT assigns each pathology to anatomically appropriate support and pools features only within that support, producing pathology-aware representations such as kidney-pooled features for renal cyst and liver–spl… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ORACLE–CT anatomy-aware support pooling. ORACLE–CT modifies the aggregation stage of multi-label CT classification while leaving the encoder architecture unchanged. Each backbone is trained separately, but within a given backbone the same full-volume encoder provides features for all labels. (A) An abdominal CT volume is paired with a multi-organ segmentation, which provides coarse anatomical s… view at source ↗
Figure 3
Figure 3. Figure 3: compares per-label prevalence across MERLIN, Duke–Abdomen, and AMOS for the harmonized 10- label subset used for three-way evaluation. Several findings showed marked prevalence differences between datasets, indicating that external transfer was assessed under genuine cross-dataset case-mix shift. Lavsen Dahal et al.: Preprint submitted to Elsevier Page 10 of 25 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: ORACLE-CT: Encoder-dependent internal test results on MERLIN (n=5,082). (A) Macro-averaged AUROC on the MERLIN test set across three pooling modes for each encoder backbone. Error bars denote 95% bootstrap confidence intervals (1,000 bootstrap replicates). (B) Per-disease ΔAUROC relative to the GAP baseline for mask-free attention pooling (blue circles) and support-masked attention pooling (orange diamonds… view at source ↗
Figure 5
Figure 5. Figure 5: External generalization on the harmonized 10-label subset. (A) Macro AUROC across MERLIN, Duke–Abdomen, and AMOS for GAP, mask-free attention pooling, and support-masked pooling across DINOv3, I3D–ResNet-121, and Pillar–0. Error bars denote 95% bootstrap confidence intervals. (B) Disease-level decomposition of the Duke-to-AMOS shift for the support-masked model on the 10 harmonized labels. Blue circles den… view at source ↗
Figure 6
Figure 6. Figure 6: Disease-level MERLIN-to-Duke transfer on the shared 27-label subset. Heatmap of per-finding ΔAUROC relative to GAP on Duke–Abdomen for mask-free attention pooling and support-masked pooling across DINOv3, I3D–ResNet-121, and Pillar–0. Blue indicates improvement over uniform full-lattice aggregation, red indicates lower AUROC than GAP, and values denote absolute AUROC differences. Rows are ordered by the av… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative audit of support-masked attention. Representative axial CT slices with anatomical support contours and support-masked attention overlays for pleural effusion, hepatic steatosis, renal cyst, and pancreatic atrophy. Columns show true-positive examples from DINOv3 support-masked and Pillar–0 support-masked models, and representative false￾positive examples from DINOv3 support-masked models. Attent… view at source ↗
read the original abstract

Abdominal CT disease classification is challenging because each scan is a large 3D volume with many possible findings, while diagnostic evidence is often confined to specific organs or anatomical compartments. Most study-level classifiers aggregate encoder features using anatomy-agnostic pooling or attention, creating a mismatch between localized disease evidence and global evidence aggregation. We propose ORACLE--CT, an encoder-agnostic anatomy-aware aggregation framework that uses multi-organ segmentation to define label-specific anatomical supports and restrict attention pooling to relevant regions. The framework supports single-organ, multi-organ union, comparative, localized, and global support strategies. We evaluate ORACLE--CT with three encoder families: DINOv3, I3D--ResNet-121, and the radiology-native Pillar--0 encoder. Models are trained end-to-end on MERLIN and evaluated internally and under frozen external transfer to Duke--Abdomen and AMOS. Compared with global average pooling, support-masked pooling improved MERLIN macro-AUROC/AUPRC from 0.838/0.638 to 0.858/0.676 for DINOv3 and from 0.829/0.617 to 0.848/0.659 for I3D--ResNet-121. On harmonized 10-label external evaluation, DINOv3 improved on Duke--Abdomen from 0.802/0.628 to 0.835/0.683 and on AMOS from 0.742/0.313 to 0.762/0.350, with similar gains for I3D--ResNet-121. For Pillar--0, most gains came from learned attention, with smaller additional benefit from anatomical masking. ORACLE--CT improves discrimination and external robustness while preserving an auditable link between predictions and anatomical evidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes ORACLE-CT, an encoder-agnostic framework for abdominal CT disease classification that defines label-specific anatomical supports via multi-organ segmentation and restricts pooling/attention to those regions using strategies such as single-organ, multi-organ union, and comparative supports. It evaluates the approach with DINOv3, I3D-ResNet-121, and Pillar-0 encoders trained on MERLIN, reporting AUROC/AUPRC gains over global average pooling both internally (e.g., DINOv3: 0.838/0.638 to 0.858/0.676) and on external transfer to Duke-Abdomen and AMOS (e.g., DINOv3 Duke: 0.802/0.628 to 0.835/0.683), while claiming improved discrimination, robustness, and auditability.

Significance. If the results hold after addressing segmentation validation, the work provides a concrete mechanism to align feature aggregation with localized anatomical evidence in volumetric CT, addressing a known mismatch in study-level classifiers. The consistent gains across encoders and the external-transfer setting are a strength; the auditable anatomical link is a secondary contribution that could aid clinical interpretability.

major comments (2)
  1. [Abstract / framework description] Abstract and framework description: The reported improvements depend on multi-organ segmentation accurately defining the supports without omitting diagnostic evidence outside segmented regions or introducing boundary errors. No Dice, Hausdorff, or other segmentation quality metrics are provided on MERLIN, Duke-Abdomen, or AMOS, and no ablation on segmentation noise is shown; this is load-bearing for the central claim that support-masked pooling yields the observed gains.
  2. [Evaluation] Evaluation: Performance deltas (e.g., DINOv3 MERLIN AUROC 0.838 o0.858) are presented as point estimates without error bars, standard deviations across runs, or statistical significance tests, making it impossible to determine whether the gains are robust or could arise from post-hoc choices in support strategy or hyperparameters.
minor comments (1)
  1. [Abstract] The abstract refers to a 'harmonized 10-label external evaluation' but provides no details on label mapping, harmonization procedure, or how the 10 labels align across datasets.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the detailed review and the recommendation for major revision. We appreciate the focus on the dependence on segmentation quality and the need for statistical rigor in the reported results. Below we provide point-by-point responses and commit to revisions that address these concerns.

read point-by-point responses

  1. Referee: [Abstract / framework description] Abstract and framework description: The reported improvements depend on multi-organ segmentation accurately defining the supports without omitting diagnostic evidence outside segmented regions or introducing boundary errors. No Dice, Hausdorff, or other segmentation quality metrics are provided on MERLIN, Duke-Abdomen, or AMOS, and no ablation on segmentation noise is shown; this is load-bearing for the central claim that support-masked pooling yields the observed gains.

    Authors: We agree that the accuracy of the multi-organ segmentation is critical to the validity of our claims. The segmentation is performed using a pre-trained model, and while we did not include quantitative metrics in the initial submission, the framework's gains are demonstrated consistently across three different encoders and in external validation settings. To strengthen the manuscript, we will add Dice coefficient and other relevant metrics for the segmentation on the MERLIN dataset in a new section or appendix, along with a sensitivity analysis to simulated segmentation errors (e.g., boundary perturbations). This will directly address the load-bearing nature of this assumption. revision: yes

  2. Referee: [Evaluation] Evaluation: Performance deltas (e.g., DINOv3 MERLIN AUROC 0.838 to 0.858) are presented as point estimates without error bars, standard deviations across runs, or statistical significance tests, making it impossible to determine whether the gains are robust or could arise from post-hoc choices in support strategy or hyperparameters.

    Authors: We concur that providing measures of variability and statistical testing would improve the interpretability of the results. In the revised version, we will report mean and standard deviation of AUROC/AUPRC over at least three independent training runs with different random seeds for the key comparisons. Additionally, we will include p-values from appropriate statistical tests (such as the DeLong test for comparing AUROCs) to establish the significance of the observed improvements. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; empirical gains are measured, not constructed by definition.

full rationale

The paper describes an empirical framework that applies pre-existing multi-organ segmentation to mask pooling regions in encoder features, then reports measured AUROC/AUPRC improvements on held-out MERLIN data and external Duke/AMOS transfers. No derivation chain, equation, or first-principles claim reduces to its own inputs by construction. No self-citations are load-bearing for uniqueness or ansatz adoption, no fitted parameters are relabeled as predictions, and no renaming of known results occurs. The segmentation accuracy is an explicit modeling assumption whose validity is external to the reported performance deltas.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based on abstract only. No free parameters explicitly mentioned. Relies on standard performance metrics and assumes availability of pre-trained segmentation and encoder models from prior literature.

axioms (1)
  • standard math Standard definitions and computation of AUROC and AUPRC
    Used to quantify and compare model performance improvements.

pith-pipeline@v0.9.1-grok · 5876 in / 1310 out tokens · 51927 ms · 2026-06-28T06:23:18.771513+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 9 linked inside Pith

  1. [1]

    Pillar-0: A new frontier for radiology foundation models

    Agrawal, K.K., Liu, L., Lian, L., Nercessian, M., Harguindeguy, N., Wu, Y., Mikhael, P., Lin, G., Sequist, L.V., Fintelmann, F., et al., 2025. Pillar-0: A new frontier for radiology foundation models. arXiv preprint arXiv:2511.17803

    arXiv 2025

  2. [2]

    Neuralmachinetranslationbyjointlylearningtoalignandtranslate

    Bahdanau,D.,Cho,K.,Bengio,Y.,2014. Neuralmachinetranslationbyjointlylearningtoalignandtranslate. arXivpreprintarXiv:1409.0473

    Pith/arXiv arXiv 2014

  3. [3]

    A pan-organ vision-language model for generalizable 3d ct representations

    Beeche, C., Kim, J., Tavolinejad, H., Zhao, B., Sharma, R., Duda, J., Gee, J., Dako, F., Verma, A., Morse, C., et al., 2025. A pan-organ vision-language model for generalizable 3d ct representations. medRxiv

    2025

  4. [4]

    Abdomengen:Sequentialvolume-conditioneddiffusionframeworkforabdominalanatomy generation

    Bhandari,Y.,Dahal,L.,Segars,P.,Lo,J.Y.,2026. Abdomengen:Sequentialvolume-conditioneddiffusionframeworkforabdominalanatomy generation. arXiv preprint arXiv:2604.12969

    Pith/arXiv arXiv 2026

  5. [5]

    Blankemeier, L., Cohen, J.P., Kumar, A., Van Veen, D., Gardezi, S.J.S., Paschali, M., Chen, Z., Delbrouck, J.B., Reis, E., Truyts, C., et al.,

  6. [6]

    Research Square , rs–3

    Merlin: A vision language foundation model for 3d computed tomography. Research Square , rs–3

  7. [7]

    Quovadis,actionrecognition?anewmodelandthekineticsdataset,in:proceedingsoftheIEEEConference on Computer Vision and Pattern Recognition, pp

    Carreira,J.,Zisserman,A.,2017. Quovadis,actionrecognition?anewmodelandthekineticsdataset,in:proceedingsoftheIEEEConference on Computer Vision and Pattern Recognition, pp. 6299–6308

    2017

  8. [8]

    Peritoneal carcinomatosis and its mimics: review of ct findings for differential diagnosis

    Cho, J.H., Kim, S.S., 2020. Peritoneal carcinomatosis and its mimics: review of ct findings for differential diagnosis. Journal of the Belgian Society of Radiology 104, 8

    2020

  9. [9]

    Foundation models for general ct image diagnosis challenge Accessed: 2026-05-02

    CodaBench, 2026. Foundation models for general ct image diagnosis challenge Accessed: 2026-05-02

    2026

  10. [10]

    Xcat 3.0: A comprehensive library of personalized digital twins derived from ct scans

    Dahal, L., Ghojoghnejad, M., Vancoillie, L., Ghosh, D., Bhandari, Y., Kim, D., Ho, F.C., Tushar, F.I., Luo, S., Lafata, K.J., et al., 2025. Xcat 3.0: A comprehensive library of personalized digital twins derived from ct scans. Medical Image Analysis , 103636

    2025

  11. [11]

    Solving the multiple instance problem with axis-parallel rectangles

    Dietterich, T.G., Lathrop, R.H., Lozano-Pérez, T., 1997. Solving the multiple instance problem with axis-parallel rectangles. Artificial intelligence 89, 31–71. Lavsen Dahal et al.:Preprint submitted to ElsevierPage 19 of 25 ORACLE-CT: Anatomy-Aware Support Pooling for CT Classification

    1997

  12. [12]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al., 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

    Pith/arXiv arXiv 2020

  13. [13]

    Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes

    Draelos, R.L., Dov, D., Mazurowski, M.A., Lo, J.Y., Henao, R., Rubin, G.D., Carin, L., 2021. Machine-learning-based multiple abnormality prediction with large-scale chest computed tomography volumes. Medical image analysis 67, 101857

    2021

  14. [14]

    Evaluating large language models for zero-shot disease labeling in ct radiology reports across organ systems

    Garcia-Alcoser, M.E., GhojoghNejad, M., Tushar, F.I., Kim, D., Lafata, K.J., Rubin, G.D., Lo, J.Y., 2025. Evaluating large language models for zero-shot disease labeling in ct radiology reports across organ systems. arXiv preprint arXiv:2506.03259

    arXiv 2025

  15. [15]

    Diagnostic accuracy of ct for the detection of hepatic steatosis: a systematic review and meta-analysis

    Haghshomar, M., Antonacci, D., Smith, A.D., Thaker, S., Miller, F.H., Borhani, A.A., 2024. Diagnostic accuracy of ct for the detection of hepatic steatosis: a systematic review and meta-analysis. Radiology 313, e241171

    2024

  16. [16]

    Developing generalist foundation models from a multimodal dataset for 3d computed tomography

    Hamamci, I.E., Er, S., Wang, C., Almas, F., Simsek, A.G., Esirgun, S.N., Doga, I., Durugol, O.F., Dai, W., Xu, M., et al., 2024. Developing generalist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834

    arXiv 2024

  17. [17]

    Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    He, K., Zhang, X., Ren, S., Sun, J., 2016. Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778

    2016

  18. [18]

    Segmentation helps understanding: Mask-infused vision-language pre-training for 3d medical images

    Hu, Y., Luo, X., Wang, Z., Li, D., Qiu, L., . Segmentation helps understanding: Mask-infused vision-language pre-training for 3d medical images

  19. [19]

    Attention-based deep multiple instance learning, in: International conference on machine learning, PMLR

    Ilse, M., Tomczak, J., Welling, M., 2018. Attention-based deep multiple instance learning, in: International conference on machine learning, PMLR. pp. 2127–2136

    2018

  20. [20]

    Le Guellec, B., Lefèvre, A., Geay, C., Shorten, L., Bruge, C., Hacein-Bey, L., Amouyel, P., Pruvo, J.P., Kuchcinski, G., Hamroun, A.,

  21. [21]

    Radiology: Artificial Intelligence 6, e230364

    Performance of an open-source large language model in extracting information from free-text radiology reports. Radiology: Artificial Intelligence 6, e230364

  22. [22]

    Unified supervision for vision-language modeling in 3d computed tomography

    Lee, H.C., Liu, Z., Ahmed, H., Kim, S., Huver, S., Nath, V., Fayad, Z.A., Deyer, T., Mei, X., 2025. Unified supervision for vision-language modeling in 3d computed tomography. arXiv preprint arXiv:2509.01554

    arXiv 2025

  23. [23]

    Greenrfm: Toward a resource-efficient radiology foundation model

    Li, Y., Ming, S., Zhao, M., Lai, H., Wang, R., Zhou, R., Wang, R., Li, Y., Wei, W., Zhou, S.K., 2026. Greenrfm: Toward a resource-efficient radiology foundation model. arXiv preprint arXiv:2603.06467

    Pith/arXiv arXiv 2026

  24. [24]

    Ct-glip: 3d grounded language-image pretraining with ct scans and radiology reports for full-body scenarios

    Lin, J., Xia, Y., Zhang, J., Yan, K., Lu, L., Luo, J., Zhang, L., 2024. Ct-glip: 3d grounded language-image pretraining with ct scans and radiology reports for full-body scenarios. arXiv preprint arXiv:2404.15272

    arXiv 2024

  25. [25]

    Network in network

    Lin, M., Chen, Q., Yan, S., 2013. Network in network. arXiv preprint arXiv:1312.4400

    Pith/arXiv arXiv 2013

  26. [26]

    Does dinov3 set a new medical vision standard? arXiv preprint arXiv:2509.06467

    Liu, C., Chen, Y., Shi, H., Lu, J., Jian, B., Pan, J., Cai, L., Wang, J., Zhang, Y., Li, J., et al., 2025. Does dinov3 set a new medical vision standard? arXiv preprint arXiv:2509.06467

    arXiv 2025

  27. [27]

    Effective approaches to attention-based neural machine translation

    Luong, M.T., Pham, H., Manning, C.D., 2015. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025

    Pith/arXiv arXiv 2015

  28. [28]

    Distributionofintra-abdominalmalignantseeding:dependencyondynamicsofflowofasciticfluid

    MEYERS,M.A.,1973. Distributionofintra-abdominalmalignantseeding:dependencyondynamicsofflowofasciticfluid. AmericanJournal of Roentgenology 119, 198–206

    1973

  29. [29]

    Feasibility of using the privacy-preserving large language model vicuna for labeling radiology reports

    Mukherjee, P., Hou, B., Lanfredi, R.B., Summers, R.M., 2023. Feasibility of using the privacy-preserving large language model vicuna for labeling radiology reports. Radiology 309, e231147

    2023

  30. [30]

    Computed tomography (ct) exams URL:https://www.oecd.org/en/ data/indicators/computed-tomography-ct-exams.html

    Organisation for Economic Co-operation and Development, 2025. Computed tomography (ct) exams URL:https://www.oecd.org/en/ data/indicators/computed-tomography-ct-exams.html. accessed: 2025-10-29

    2025

  31. [31]

    Post-contrast ct liver attenuation alone is superior to the liver–spleen difference for identifying moderate hepatic steatosis

    Pickhardt, P.J., Blake, G.M., Moeller, A., Garrett, J.W., Summers, R.M., 2024. Post-contrast ct liver attenuation alone is superior to the liver–spleen difference for identifying moderate hepatic steatosis. European Radiology 34, 7041–7052

    2024

  32. [32]

    Imaging information overload: Quantifying the burden of interpretive and non-interpretive tasks for computed tomography angiography for aortic pathologies in emergency radiology

    Pourvaziri,A.,Narayan,A.K.,Tso,D.,Baliyan,V.,GloverIV,M.,Bizzo,B.C.,Kako,B.,Succi,M.D.,Lev,M.H.,Flores,E.J.,2022. Imaging information overload: Quantifying the burden of interpretive and non-interpretive tasks for computed tomography angiography for aortic pathologies in emergency radiology. Current Problems in Diagnostic Radiology 51, 546–551

    2022

  33. [33]

    Learning transferable visual models from natural language supervision, in: International conference on machine learning, PmLR

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al., 2021. Learning transferable visual models from natural language supervision, in: International conference on machine learning, PmLR. pp. 8748–8763

    2021

  34. [34]

    A scoping review of large language model based approaches for information extraction from radiology reports

    Reichenpfader, D., Müller, H., Denecke, K., 2024. A scoping review of large language model based approaches for information extraction from radiology reports. npj Digital Medicine 7, 222

    2024

  35. [35]

    Medgemma technical report

    Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., Xu, S., Jamil, F., Hughes, C., Lau, C., et al., 2025. Medgemma technical report. arXiv preprint arXiv:2507.05201

    Pith/arXiv arXiv 2025

  36. [36]

    Large-scale and fine-grained vision-language pre-training for enhanced ct image understanding

    Shui, Z., Zhang, J., Cao, W., Wang, S., Guo, R., Lu, L., Yang, L., Ye, X., Liang, T., Zhang, Q., et al., 2025. Large-scale and fine-grained vision-language pre-training for enhanced ct image understanding. arXiv preprint arXiv:2501.14548

    arXiv 2025

  37. [37]

    syngo.via — get the full picture

    Siemens Healthineers, . syngo.via — get the full picture. URL:https://www.siemens-healthineers.com/en-us/ molecular-imaging/pet-ct/syngo-via. siemens Healthineers USA

  38. [38]

    Bosniak classification of cystic renal masses, version 2019: an update proposal and needs assessment

    Silverman, S.G., Pedrosa, I., Ellis, J.H., Hindman, N.M., Schieda, N., Smith, A.D., Remer, E.M., Shinagare, A.B., Curci, N.E., Raman, S.S., et al., 2019. Bosniak classification of cystic renal masses, version 2019: an update proposal and needs assessment. Radiology 292, 475–488

    2019

  39. [39]

    Siméoni, O.,Vo, H.V.,Seitzer, M.,Baldassarre, F., Oquab,M., Jose,C., Khalidov,V., Szafraniec, M.,Yi, S.,Ramamonjisoa, M.,et al.,2025. Dinov3. arXiv preprint arXiv:2508.10104

    Pith/arXiv arXiv 2025

  40. [40]

    Projectedlifetimecancerrisksfromcurrentcomputedtomographyimaging

    Smith-Bindman, R., Chu, P.W., Firdaus, H.A., Stewart, C., Malekhedayat, M., Alber, S., Bolch, W.E., Mahendra, M., de González, A.B., Miglioretti,D.L.,2025. Projectedlifetimecancerrisksfromcurrentcomputedtomographyimaging. JAMAinternalmedicine185,710–719

    2025

  41. [41]

    Simpleware auto segmentation tools

    Synopsys, . Simpleware auto segmentation tools. URL:https://www.synopsys.com/simpleware/software/auto-segmentation. html. synopsys Simpleware

  42. [42]

    Classificationofmultiplediseases on body ct scans using weakly supervised deep learning

    Tushar,F.I.,D’Anniballe,V.M.,Hou,R.,Mazurowski,M.A.,Fu,W.,Samei,E.,Rubin,G.D.,Lo,J.Y.,2021. Classificationofmultiplediseases on body ct scans using weakly supervised deep learning. Radiology: Artificial Intelligence 4, e210026

    2021

  43. [43]

    Attention is all you need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Advances in neural information processing systems 30. Lavsen Dahal et al.:Preprint submitted to ElsevierPage 20 of 25 ORACLE-CT: Anatomy-Aware Support Pooling for CT Classification

    2017

  44. [44]

    Quantifying radiology resident fatigue: analysis of preliminary reports

    Vosshenrich, J., Brantner, P., Cyriac, J., Boll, D.T., Merkle, E.M., Heye, T., 2021. Quantifying radiology resident fatigue: analysis of preliminary reports. Radiology 298, 632–639

    2021

  45. [45]

    Wald, T., Hamamci, I.E., Gao, Y., Bond-Taylor, S., Sharma, H., Ilse, M., Lo, C., Melnichenko, O., Codella, N.C., Wetscherek, M.T., et al.,

  46. [46]

    arXiv preprint arXiv:2510.15042

    Comprehensive language-image pre-training for 3d medical image understanding. arXiv preprint arXiv:2510.15042

    arXiv

  47. [47]

    Non-local neural networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Wang, X., Girshick, R., Gupta, A., He, K., 2018. Non-local neural networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 7794–7803

    2018

  48. [48]

    Totalsegmentator: robust segmentation of 104 anatomic structures in ct images

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al., 2023. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5, e230024

    2023

  49. [49]

    Show, attend and tell: Neural image caption generation with visual attention, in: International conference on machine learning, PMLR

    Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., Bengio, Y., 2015. Show, attend and tell: Neural image caption generation with visual attention, in: International conference on machine learning, PMLR. pp. 2048–2057

    2015

  50. [50]

    Qwen3 technical report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al., 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    Pith/arXiv arXiv 2025

  51. [51]

    Deep sets

    Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdinov, R.R., Smola, A.J., 2017. Deep sets. Advances in neural information processing systems 30

    2017

  52. [52]

    Metrics are computed over the full 30-label MERLIN test set

    Supplementary Material Lavsen Dahal et al.:Preprint submitted to ElsevierPage 21 of 25 ORACLE-CT: Anatomy-Aware Support Pooling for CT Classification Table S7 MERLIN full 30-label performance with 95% bootstrap confidence intervals.Values are macro-averaged point estimates with 95% bootstrap confidence intervals in brackets. Metrics are computed over the ...