Pith. sign in

REVIEW 3 major objections 6 minor 52 references

Content-based 3D Image Retrieval and a ColBERT-inspired Re-ranking for Tumor Flagging and Staging

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that adapting ColBERT's late interaction to whole 3D volumes lets a content-based retrieval system flag colon and lung tumors as accurately without organ segmentation as with it.

desk verdict A genuinely novel volumetric re-ranking method with solid colon flagging results, but the headline lung significance and the 'best baseline' comparisons need a careful statistical rework. read the letter →

arxiv 2507.17412 v2 pith:6DD6SLTL submitted 2025-07-23 cs.CV cs.AIcs.IR

Farnaz Khun Jush , Steffen Vogler , Matthias Lenga This is my paper
classification cs.CVcs.AIcs.IR
keywords content-basedimageretrieval3Dmedicalimagingre-rankingColBERTlateinteractiontumorflaggingstagingcomputedtomographysegmentation-free
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a content-based retrieval system for 3D radiology volumes can work directly on raw, unsegmented CT scans if it borrows ColBERT's late-interaction idea: compare every slice of the query volume against every slice of a candidate volume, then sum the best matches. This re-ranking pass, called C-MIR, is applied to the top 20 volumes returned by a fast slice-level search. Across four tumor sites and three pretrained feature extractors, C-MIR improves tumor flagging over the prior aggregation baselines, with statistically significant gains for colon and lung tumors (best average precision 0.807 and 0.942 with the self-supervised ensemble embeddings), and its scores are the same whether or not the database was filtered by organ segmentation. The practical point is that a hospital archive resembling a PACS could be searched without paying for segmentation or organ-specific preprocessing, while staging retrieved cases remains harder and more variable.

What carries the argument

The central object is the volume-to-volume embedding-matrix similarity: for a query volume $V_Q=[q_1,\dots,q_n]$ and a candidate volume $V_k=[v_1,\dots,v_m]$, C-MIR builds the cosine-similarity matrix $\operatorname{SIM}(M_{V_Q}, M_{V_k})$ and assigns the rank score $\operatorname{RS}(V_k)=\sum_{i=1}^{n}\max_{j=1,\dots,m}\operatorname{SIM}_{i,j}$. This is the ColBERT-style late interaction: each query slice attends to the best matching slice in the candidate volume, and the sum pools those matches across the whole query. The method is applied only to the top $M=20$ volumes from the initial slice-level search, so the computation scales with the candidate count, not the database size. It is this matrix operation that the paper credits with implicitly localizing tumor regions, since the maximum over candidate slices can land on informative slices even when the database contains many background slices.

What would settle it

Measure the initial slice-level retrieval's recall@20 on a held-out set of tumor queries before applying C-MIR. Because C-MIR only reorders the top-20 candidate volumes, any query whose correct volume is not in that shortlist cannot be rescued, so showing that such queries exist and that C-MIR's improvements disappear when recall@20 is low would settle that the method's value is bounded by the first-stage search.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the late-interaction mechanism, originally built for text passages, transfers to volumetric images: treating slices as tokens and volumes as passages lets a retrieval system re-rank candidate volumes by their full 3D context. Concretely, C-MIR computes a similarity matrix between the query volume's slice embeddings and each candidate volume's slice embeddings, takes the row-wise maximum over candidate slices, and sums these maxima to form a rank score. The paper reports that this re-ranking is consistently the best method for colon tumor flagging across all feature extractors and database configurations, and best for lung flagging with the self-supervised ensemble embeddings, while producing identical average precision with and without segmentation masks, which is the evidence for the claim that C-MIR implicitly localizes the region of interest rather than needing a pre-segmentation step.

Load-bearing premise

The re-ranking is limited to the top 20 volumes returned by the initial slice-level retrieval, so if the relevant volume is not in that shortlist, C-MIR cannot bring it into the re-ranked results.

Editorial extensions

If this is right

  • For colon and lung tumor flagging, C-MIR's average precision is statistically significantly higher than the best baseline (p<0.05), with the largest gains on organ-specific databases using the self-supervised ensemble embeddings.
  • Because C-MIR's results are identical with and without segmentation masks, a deployment could drop the segmentation-enrichment step entirely and still match the accuracy of a search space filtered to the organ.
  • The re-ranking cost grows with the number of top-k candidates rather than the whole database, so the method stays tractable for large archives, with roughly 6.14B FLOPs and under 15MB of GPU memory for a 300-slice query re-ranking 20 volumes.
  • For tumor staging, C-MIR is the best method for all organs in the organ-agnostic database, but staging gains are variable, so the paper treats staging as an open problem needing higher-resolution or scale-aware features.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The max-over-slices operation in the rank score acts like a soft, learned localization: because C-MIR's scores are identical with and without segmentation masks, one can infer that the late-interaction similarity is already concentrating on tumor-bearing slices; a direct test would be to visualize which slices receive the maximum similarity.
  • Because C-MIR treats any volume as a set of slice embeddings, the same re-ranking could apply to other variable-length volumetric tasks, such as lesion tracking across time points or retrieval of non-tumor findings, without changing the method.
  • The staging results suggest that slice-level embeddings discard absolute scale and pixel-spacing information; one testable extension is to include physical spacing metadata in the rank score to recover stage-relevant size cues.
  • A practical safeguard is that if initial-retrieval recall@20 is the binding constraint, combining C-MIR with a broader first-stage candidate set, such as the top 50 volumes, should improve final flagging at modest extra cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The manuscript introduces C-MIR, a re-ranking method for 3D medical content-based image retrieval that adapts ColBERT's late interaction mechanism to volumetric data. After an initial slice-level retrieval, the top-20 candidate volumes are re-scored by summing, over query slices, the maximum cosine similarity between the query slice and all slices of each candidate volume. The authors evaluate tumor flagging and staging on four MSD organs (colon, liver, lung, pancreas) using three feature extractors (DreamSim, SwinTransformer, BioMedClip) and three database configurations (organ-specific with/without segmentation, and organ-agnostic), repeating each experiment 10 times with different seeds. They report that C-MIR improves tumor flagging, particularly for colon and lung (p<0.05), and claim that C-MIR effectively localizes regions of interest, thereby eliminating the need for pre-segmentation.

Significance. If the central claims hold, the paper would be a useful practical contribution to medical CBIR: it offers a segmentation-free re-ranking method with a principled late-interaction design, a broad evaluation (three extractors, three database configurations, four organs), and a reasonable computational cost analysis in Section 4.4. The 10-seed volume-level split protocol and the use of two-sided Wilcoxon tests are appropriate, and the paper is transparent about the dependence of re-ranking on the initial top-20 retrieval (Section 4.3). However, the statistical reporting contains inconsistencies that affect the headline significance claims, and the localization claim currently overreaches the evidence. The staging results are also based on an automatic surrogate for TNM staging, which limits their clinical interpretation.

major comments (3)
  1. [Section 3.3 (Tables 7-8 vs Tables 4-5)] The claim that the two-sided Wilcoxon test compares C-MIR against 'the best method for each database' is not consistently implemented. For colon/DreamSim with segmentation, Table 4 shows Max-Score AP 0.782, above RRF's 0.771, yet Table 7 lists RRF as the comparator (p=0.003); without segmentation, Max-Score AP 0.792 again exceeds RRF 0.777, and RRF is again listed (p=0.048). For pancreas/BioMedClip in the organ-agnostic database, Count-Base AP 0.820 exceeds RRF 0.815, while Table 7 lists RRF (p=0.001). The reported p-values therefore cannot all be interpreted as 'C-MIR versus the best-performing method,' and the headline colon/lung significance partly relies on comparisons against a weaker baseline. Please recompute the Wilcoxon tests against the true best baseline for every cell and revise Tables 7-8 and the surrounding text.
  2. [Section 3.3.1 (Table 7) and Abstract] The abstract's lung claim 'p<0.05' rests on two p-values, 0.037 and 0.049, from the DreamSim/organ-specific cells in Table 7, computed from 10 seeds with no correction for the 36 flagging comparisons in Table 7 (and the additional 36 staging comparisons in Table 8). A p-value of 0.049 would not survive a Benjamini-Hochberg correction at q=0.05, and the comparator-selection inconsistency in the previous comment makes it even less reliable. The authors should report the total number of tests performed, apply a suitable multiplicity correction, or limit the significance claim to the robust colon results (e.g., BioMedClip and SwinTransformer cells with p<=0.005).
  3. [Abstract and Section 4.1] The claim that C-MIR 'effectively localize[s] the region of interest, eliminating the need for pre-segmentation' is an overinterpretation. C-MIR re-ranks entire volumes using full embedding matrices and never uses segmentation information; the observation that AP is identical with and without segmentation is a by-construction property of the method's inputs, not empirical evidence of spatial localization. No localization metric (e.g., overlap of the implicitly attended slices with the tumor/organ mask) is reported. The authors should either drop the localization wording or add a dedicated evaluation that measures whether the method identifies the relevant slices within the volume.
minor comments (6)
  1. [Section 3.2.2] The pancreas staging paragraph says performance 'drops for DreamSim and SwinTransformer embeddings,' but the preceding sentence names SwinTransformer as a case where C-MIR is best and Table 5 shows C-MIR matching Sum-Sim for DreamSim; please correct this contradiction.
  2. [Table 5] The heading contains the typo 'Stagging'; it should read 'Staging.'
  3. [Section 2.7.2] The text says 'hightest RRF score'; change to 'highest.'
  4. [Table B1] The column header 'Max-Sim' is inconsistent with 'Max-Score' used in Tables 4-6; unify the terminology.
  5. [Section 2.4.1] The sentence 'sampling is performed with replacement' is ambiguous: if volumes are sampled with replacement, a volume could appear in both query and database across seeds, and the volume-level split guarantee needs clarification. Please specify the exact sampling procedure.
  6. [Data Availability] For a benchmark paper, providing the exact split indices, seed values, and model versions as a supplementary file would markedly improve reproducibility; 'available upon request' is insufficient for the quantitative claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the C-MIR re-ranking result is an independent empirical claim, and the segmentation-free property is architectural rather than a fitted prediction.

full rationale

The paper's central claim is that C-MIR, a ColBERT-inspired late-interaction re-ranker over full-volume embedding matrices, improves tumor flagging AP and is invariant to segmentation-based slice pre-filtering. This claim is supported by measured mean AP values across 10 seeds (Tables 4 and 5) and by Wilcoxon signed-rank tests (Tables 7 and 8). C-MIR's rank score in Eq. (4) is defined from embedding matrices and max-pooled cosine similarities; no parameter is fitted to the flagging labels and no prediction is derived from a fitted value. The 'segmentation-free' property is a direct architectural consequence of operating on all slices, but the paper does not use that definition to derive the reported AP gains; the AP invariance is an empirical observation. The Count-base, Max-Score, and Sum-Sim aggregation baselines are taken from prior work, some of it the authors' own ([19], [20], [35]), but they serve as baselines rather than as the source of the claimed advantage. The ColBERT inspiration is explicitly attributed to an external source [13]. Self-citations concerning HNSW indexing and Count-base aggregation are not load-bearing for the core flagging result. The acknowledged limitation that re-ranking depends on initial retrieval quality (Section 4.3) is a scope condition, not a circular step. Statistical concerns about baseline selection and multiple testing are correctness risks, not circularity. No step in the derivation reduces to its own input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that pretrained 2D embeddings encode tumor-relevant features, that MSD ground-truth masks and [19]'s derived staging are valid labels, and that the initial top-20 retrieval contains the correct volume. No new entities are introduced; the only free parameters are evaluation choices (sampling proportion p=25%, RRF k=60, top-k M=20).

free parameters (3)
  • Positive sampling proportion p = 25%
    Used in Section 2.4.1/2.4.2 to select positive cases per tumor stage; chosen by the authors and affects query/database composition.
  • RRF smoothing parameter k = 60
    Set following [48, 51] in Section 2.7.2; a standard choice but a free constant in the fusion baseline.
  • Top-k candidate volume count M = 20
    Number of initially retrieved volumes re-ranked by C-MIR, following [19] (Section 2.6); the re-ranking cannot recover relevant volumes outside this list.
assumptions (4)
  • domain assumption Pretrained 2D slice embeddings are sufficiently discriminative for tumor presence when aggregated over volumes.
    The entire retrieval pipeline in Sections 2.2-2.6 assumes cosine similarity in the embedding space captures clinical relevance.
  • domain assumption MSD tumor masks and [19]'s TNM-derived staging labels are valid ground truth for flagging and staging.
    Used in Section 2.3 and throughout evaluation; staging labels derive from automatic segmentation and morphological rules, not clinical histopathology.
  • standard math Standard linear algebra operations (cosine similarity, max-pooling, summation) are well-defined and correctly implemented.
    Equations (1)-(4) in Sections 2.6-2.7.1 assume standard properties of Euclidean space.
  • domain assumption HNSW approximate nearest neighbor search retrieves the true top-k nearest slices with negligible error.
    Section 2.1 uses FAISS HNSWFlat without error analysis; approximate search could affect the initial retrieval list.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Content-based 3D Image Retrieval and a ColBERT-inspired Re-ranking for Tumor Flagging and Staging." pith.science (2026). https://pith.science/paper/6DD6SLTL

@misc{pith2026250717412,
  author       = {Pith},
  title        = {Pith review of: Content-based 3D Image Retrieval and a ColBERT-inspired Re-ranking for Tumor Flagging and Staging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6DD6SLTL}},
  note         = {Machine review of arXiv:2507.17412}
}
read the original abstract

The increasing volume of medical images poses challenges for radiologists in retrieving relevant cases. Content-based image retrieval (CBIR) systems offer potential for efficient access to similar cases, yet lack standardized evaluation and comprehensive studies. Building on prior studies for tumor characterization via CBIR, this study advances CBIR research for volumetric medical images through three key contributions: (1) a framework eliminating reliance on pre-segmented data and organ-specific datasets, aligning with large and unstructured image archiving systems, i.e. PACS in clinical practice; (2) introduction of C-MIR, a novel volumetric re-ranking method adapting ColBERT's contextualized late interaction mechanism for 3D medical imaging; (3) comprehensive evaluation across four tumor sites using three feature extractors and three database configurations. Our evaluations highlight the significant advantages of C-MIR. We demonstrate the successful adaptation of the late interaction principle to volumetric medical images, enabling effective context-aware re-ranking. A key finding is C-MIR's ability to effectively localize the region of interest, eliminating the need for pre-segmentation of datasets and offering a computationally efficient alternative to systems relying on expensive data enrichment steps. C-MIR demonstrates promising improvements in tumor flagging, achieving improved performance, particularly for colon and lung tumors (p<0.05). C-MIR also shows potential for improving tumor staging, warranting further exploration of its capabilities. Ultimately, our work seeks to bridge the gap between advanced retrieval techniques and their practical applications in healthcare, paving the way for improved diagnostic processes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 43 canonical work pages

  1. [19]

    arXiv preprint arXiv:2311.13752 (2023) 30

    Abacha, A.B., Santamaria-Pang, A., Lee, H.H., Merkow, J., Cai, Q., Devarakonda, S.T., Islam, A., Gong, J., Lungren, M.P., Lin, T., et al.: 3d-mir: A benchmark and empirical study on 3d medical image retrieval in radiology. arXiv preprint arXiv:2311.13752 (2023) 30

    arXiv 2023

  2. [1]

    IEEE Transactions on Circuits and Systems for Video Technology 32(5), 2687–2704 (2021)

    Dubey, S.R.: A decade survey of content based image retrieval using deep learn- ing. IEEE Transactions on Circuits and Systems for Video Technology 32(5), 2687–2704 (2021)

    work page 2021

  3. [2]

    Multimedia Tools and Applications 81(12), 16621–16644 (2022)

    Wang, W., Jiao, P., Liu, H., Ma, X., Shang, Z.: Two-stage content based image retrieval using sparse representation and feature fusion. Multimedia Tools and Applications 81(12), 16621–16644 (2022)

    work page 2022

  4. [3]

    Journal of Critical Reviews 7(11), 712–721 (2020)

    Moirangthem, M., Singh, T.: Content based medical image retrieval (cbmir): A survey of region of interest (roi) and perceptual hash values. Journal of Critical Reviews 7(11), 712–721 (2020)

    work page 2020

  5. [4]

    Neurocomputing 266, 8–20 (2017)

    Qayyum, A., Anwar, S.M., Awais, M., Majid, M.: Medical image retrieval using deep convolutional neural network. Neurocomputing 266, 8–20 (2017)

    work page 2017

  6. [5]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Hu, B., Vasu, B., Hoogs, A.: X-mir: Explainable medical image retrieval. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 440–450 (2022)

    work page 2022

  7. [6]

    Cogent Engineering 8(1), 1927469 (2021)

    Hameed, I.M., Abdulhussain, S.H., Mahmmod, B.M.: Content-based image retrieval: A review of recent trends. Cogent Engineering 8(1), 1927469 (2021)

    work page 2021

  8. [7]

    Journal of digital imaging 24, 208–222 (2011)

    Akg¨ ul, C.B., Rubin, D.L., Napel, S., Beaulieu, C.F., Greenspan, H., Acar, B.: Content-based image retrieval in radiology: current status and future directions. Journal of digital imaging 24, 208–222 (2011)

    work page 2011

Show all 52 references
  1. [8]

    IEEE Access 13, 68066–68083 (2025)

    Khun Jush, F., Vogler, S., Truong, T., Lenga, M.: Content-based image retrieval for multi-class volumetric radiology images: A benchmark study. IEEE Access 13, 68066–68083 (2025)

    2025

  2. [9]

    Washington DC: Human Health Services (2012)

    Portability, I., Act, A.: Guidance regarding methods for de-identification of pro- tected health information in accordance with the health insurance portability and 29 accountability act (hipaa) privacy rule. Washington DC: Human Health Services (2012)

    2012

  3. [10]

    Springer, 1–26 (2003)

    Long, F., Zhang, H., Feng, D.D.: Fundamentals of content-based image retrieval, in multimedia information retrieval and management–technological fundamentals and applications. Springer, 1–26 (2003)

    2003

  4. [11]

    Scientific reports 12(1), 20732 (2022)

    Silva, W., Gon¸ calves, T., H¨ arm¨ a, K., Schr¨ oder, E., Obmann, V.C., Barroso, M.C., Poellinger, A., Reyes, M., Cardoso, J.S.: Computer-aided diagnosis through medical image retrieval in radiology. Scientific reports 12(1), 20732 (2022)

    2022

  5. [12]

    IEEE Access 8, 194541–194550 (2020)

    Ahmed, A., Malebary, S.J.: Query expansion based on top-ranked images for content-based medical image retrieval. IEEE Access 8, 194541–194550 (2020)

    2020

  6. [13]

    In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

    Khattab, O., Zaharia, M.: Colbert: Efficient and effective passage search via con- textualized late interaction over bert. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 39–48 (2020)

    2020

  7. [14]

    Journal of the Association for Information Science and Technology 69(9), 1095–1108 (2018)

    Ayadi, H., Torjmen-Khemakhem, M., Daoud, M., Huang, J.X., Ben Jemaa, M.: Mf-re-rank: A modality feature-based re-ranking model for medical image retrieval. Journal of the Association for Information Science and Technology 69(9), 1095–1108 (2018)

    2018

  8. [15]

    Computers and Electrical Engineering 104, 108450 (2022)

    Vishraj, R., Gupta, S., Singh, S.: A comprehensive review of content-based image retrieval systems using deep learning and hand-crafted features in medi- cal imaging: Research challenges and future directions. Computers and Electrical Engineering 104, 108450 (2022)

    2022

  9. [16]

    Diagnostics 11(8), 1470 (2021)

    Sotomayor, C.G., Mendoza, M., Casta˜ neda, V., Far´ ıas, H., Molina, G., Pereira, G., H¨ artel, S., Solar, M., Araya, M.: Content-based medical image retrieval and intelligent interactive visual browser for medical education, research and care. Diagnostics 11(8), 1470 (2021)

    2021

  10. [17]

    arXiv preprint arXiv:2403.06567 (2024)

    Denner, S., Zimmerer, D., Bounias, D., Bujotzek, M., Xiao, S., Kausch, L., Schader, P., Penzkofer, T., J¨ ager, P.F., Maier-Hein, K.: Leveraging foundation models for content-based medical image retrieval in radiology. arXiv preprint arXiv:2403.06567 (2024)

    2024

  11. [18]

    Physics in Medicine & Biology 69(4), 045004 (2024)

    Lo, C.-M., Wang, C.-C., Hung, P.-H.: Interactive content-based image retrieval with deep learning for ct abdominal organ recognition. Physics in Medicine & Biology 69(4), 045004 (2024)

    2024

  12. [21]

    In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pp

    Truong, T., Khun Jush, F., Lenga, M.: Benchmarking pretrained vision embed- dings for near- and duplicate detection in medical images. In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1–5 (2024)

    2024

  13. [22]

    Engineering Applications of Artificial Intelligence 150, 110571 (2025)

    Mahbod, A., Saeidi, N., Hatamikia, S., Woitek, R.: Evaluating pre-trained con- volutional neural networks and foundation models as feature extractors for content-based medical image retrieval. Engineering Applications of Artificial Intelligence 150, 110571 (2025)

    2025

  14. [23]

    : The medical segmentation decathlon

    Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Land- man, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al. : The medical segmentation decathlon. Nature communications 13(1), 4128 (2022)

    2022

  15. [24]

    Radiology: Artificial Intelligence 5(5) (2023)

    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.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5) (2023)

    2023

  16. [25]

    Uicc International Union Against Cancer

    Sobin, L.H., Gospodarowicz, M.K., Wittekind, C.: TNM Classification of Malig- nant Tumours. Uicc International Union Against Cancer. John Wiley & Sons, Hoboken, NJ, USA (2009). https://books.google.de/books?id=sUaevQ0I 8kC

    2009

  17. [26]

    ACM Computing Surveys (CSUR) 46(3), 1–38 (2014)

    Mei, T., Rui, Y., Li, S., Tian, Q.: Multimedia search reranking: A literature survey. ACM Computing Surveys (CSUR) 46(3), 1–38 (2014)

    2014

  18. [27]

    Information Sciences 265, 91–104 (2014)

    Pedronette, D.C.G., Almeida, J., Torres, R.d.S.: A scalable re-ranking method for content-based image retrieval. Information Sciences 265, 91–104 (2014)

    2014

  19. [28]

    arXiv preprint arXiv:2012.07620 (2020)

    Zhang, X., Jiang, M., Zheng, Z., Tan, X., Ding, E., Yang, Y.: Understanding image retrieval re-ranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620 (2020)

    2020 arXiv

  20. [29]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Zhong, Z., Zheng, L., Cao, D., Li, S.: Re-ranking person re-identification with k- reciprocal encoding. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1318–1327 (2017)

    2017

  21. [30]

    In: 2007 IEEE 11th International Conference on Computer Vision, pp

    Chum, O., Philbin, J., Sivic, J., Isard, M., Zisserman, A.: Total recall: Automatic query expansion with a generative feature model for object retrieval. In: 2007 IEEE 11th International Conference on Computer Vision, pp. 1–8 (2007). IEEE

    2007

  22. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Lee, S., Seong, H., Lee, S., Kim, E.: Correlation verification for image retrieval. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5374–5384 (2022) 31

    2022

  23. [32]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Tan, F., Yuan, J., Ordonez, V.: Instance-level image retrieval using reranking transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 12105–12115 (2021)

    2021

  24. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Shao, S., Chen, K., Karpur, A., Cui, Q., Araujo, A., Cao, B.: Global features are all you need for image retrieval and reranking. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 11036–11046 (2023)

    2023

  25. [34]

    arXiv preprint arXiv:1810.04805 (2018)

    Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  26. [35]

    In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pp

    Khun Jush, F., Truong, T., Vogler, S., Lenga, M.: Medical image retrieval using pretrained embeddings. In: 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pp. 1–5 (2024)

    2024

  27. [36]

    IEEE transactions on pattern analysis and machine intelligence 42(4), 824–836 (2018)

    Malkov, Y.A., Yashunin, D.A.: Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE transactions on pattern analysis and machine intelligence 42(4), 824–836 (2018)

    2018

  28. [37]

    IEEE Transactions on Big Data 7(3), 535–547 (2019)

    Johnson, J., Douze, M., J´ egou, H.: Billion-scale similarity search with gpus. IEEE Transactions on Big Data 7(3), 535–547 (2019)

    2019

  29. [38]

    In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255 (2009)

    2009

  30. [39]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin trans- former: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10012–10022 (2021)

    2021

  31. [40]

    : Radimagenet: an open radiologic deep learning research dataset for effective transfer learning

    Mei, X., Liu, Z., Robson, P.M., Marinelli, B., Huang, M., Doshi, A., Jacobi, A., Cao, C., Link, K.E., Yang, T., et al. : Radimagenet: an open radiologic deep learning research dataset for effective transfer learning. Radiology: Artificial Intelligence 4(5), 210315 (2022)

    2022

  32. [41]

    arXiv preprint arXiv:2303.00915 (2023)

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023)

    2023 arXiv

  33. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Caron, M., Touvron, H., Misra, I., J´ egou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9650–9660 (2021) 32

    2021

  34. [43]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fer- nandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  35. [44]

    arXiv preprint arXiv:2306.09344 (2023)

    Fu, S., Tamir, N., Sundaram, S., Chai, L., Zhang, R., Dekel, T., Isola, P.: Dream- sim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344 (2023)

    2023 arXiv

  36. [45]

    : Learning transferable visual mod- els from natural language supervision

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. : Learning transferable visual mod- els from natural language supervision. In: International Conference on Machine Learning, pp. 8748–8763 (2021). PmLR

    2021

  37. [46]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for con- trastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp...

    2023

  38. [47]

    ACM Transactions on Information Systems 42(1), 1–35 (2023)

    Bruch, S., Gai, S., Ingber, A.: An analysis of fusion functions for hybrid retrieval. ACM Transactions on Information Systems 42(1), 1–35 (2023)

    2023

  39. [48]

    In: Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp

    Cormack, G.V., Clarke, C.L., Buettcher, S.: Reciprocal rank fusion outper- forms condorcet and individual rank learning methods. In: Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 758–759 (2009)

    2009

  40. [49]

    Nature communications 15(1), 9074 (2024)

    Jin, Q., Wang, Z., Floudas, C.S., Chen, F., Gong, C., Bracken-Clarke, D., Xue, E., Yang, Y., Sun, J., Lu, Z.: Matching patients to clinical trials with large language models. Nature communications 15(1), 9074 (2024)

    2024

  41. [50]

    arXiv preprint arXiv:2402.03367 (2024)

    Rackauckas, Z.: Rag-fusion: a new take on retrieval-augmented generation. arXiv preprint arXiv:2402.03367 (2024)

    2024 arXiv

  42. [51]

    In: European Conference on Information Retrieval, pp

    Chen, T., Zhang, M., Lu, J., Bendersky, M., Najork, M.: Out-of-domain semantics to the rescue! zero-shot hybrid retrieval models. In: European Conference on Information Retrieval, pp. 95–110 (2022). Springer

    2022

  43. [52]

    National Institute of Informatics, Tokyo, Japan (2005)

    Kishida, K.: Property of Average Precision and Its Generalization: An Exami- nation of Evaluation Indicator for Information Retrieval Experiments. National Institute of Informatics, Tokyo, Japan (2005)

    2005

  44. [53]

    International journal of computer science & information technology 4(3), 35 (2012) 33

    Zuva, K., Zuva, T.: Evaluation of information retrieval systems. International journal of computer science & information technology 4(3), 35 (2012) 33

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.