Pith. sign in

REVIEW 5 major objections 6 minor 36 references

LRMR: LLM-Driven Relational Multi-node Ranking for Lymph Node Metastasis Assessment in Rectal Cancer

T0 review · 5 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 A two-stage LLM design—first converting all lymph-node MRI patches for a patient into a ten-feature structured report, then pairwise-ranking patients on those reports—predicts rectal cancer nodal metastasis with an AUC of 0.7917 on the…

desk verdict A clearly written pilot of a plausible two-stage LLM ranking framework whose headline outperformance claim rests on a 36-patient test set and a stochastic, unrepeatable comparison stage. read the letter →

arxiv 2507.11457 v1 pith:C6F44A6C submitted 2025-07-15 cs.LG eess.IV

Yaoxian Dong , Yifan Gao , Haoyue Li , Yanfen Cui , Xin Gao This is my paper
classification cs.LGeess.IV
keywords rectalcancerlymphnodemetastasislargelanguagemodelsmultimodalLLMstructuredreportingrelationalrankingpatient-levelMRIassessmentinterpretablemedicalAI
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

LRMR (LLM-Driven Relational Multi-node Ranking) reframes rectal cancer lymph node staging as a two-stage reasoning problem. In Stage 1, a multimodal LLM (a large language model that can read images as well as text) views a composite montage of all lymph node patches from a patient and emits a structured JSON report (machine-readable text) on ten radiological features, naming which nodes show each abnormality. In Stage 2, a text-only LLM compares these reports in random pairs across patients, and the wins minus losses give each patient a relative risk score. The paper reports that on a held-out test set of 36 patients LRMR reaches an AUC (area under the receiver operating characteristic curve) of 0.7917 and an F1 of 0.7200, beating the best deep learning baseline ResNet50 (AUC 0.7708). Its two ablations—dropping the relational ranking stage (AUC 0.6875) and dropping the structured prompt (AUC 0.6458)—support the claim that both stages carry the performance.

What carries the argument

The mechanism that carries the argument is the structured JSON report produced in Stage 1 (a machine-readable text report with a fixed ten-feature checklist applied to a patient-level montage, with node identifiers attached to abnormal findings). This report is the only information passed to Stage 2, which is what decouples perception from reasoning. The second mechanism is pairwise relational ranking with a net-wins aggregate: each patient's report is compared with six randomly sampled opponents, the text LLM returns 'A', 'B', or 'Comparable,' and wins minus losses produces a single relative-risk score on which the final classification threshold is set. Together these turn a pixel-classification problem into a text-comparison problem, and they give the framework its interpretability: each comparison comes with an explicit textual explanation.

What would settle it

Compare Stage 1's JSON feature labels (for example, presence of fatty hilum, necrosis, or irregular border) against a radiologist's independent annotation of the same montages; if agreement is near chance while LRMR's test AUC stays near 0.79, the reported accuracy cannot be coming from faithful feature extraction. A complementary check is to feed radiologist-written structured reports through Stage 2 instead of LLM reports, since an unchanged or lower AUC would show the second stage is not genuinely using the feature content.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that lymph node metastasis can be assessed by making an LLM produce a reproducible, structured description of the whole set of a patient's nodes and then having a second LLM reason over those descriptions relative to other patients. Stage 1 resizes every node patch to 128x128, places them in one labeled montage, and asks the multimodal model ten clinically anchored questions covering shape, border contour, border clarity, internal signal texture, fatty hilum, necrosis, and related features, with instructions to cite node IDs for abnormal findings and return a JSON object. Stage 2 inserts two patients' JSON reports into a text-only prompt, asks for A/B/Comparable judgments with textual justification, samples six opponents per patient, and aggregates wins minus losses into a 'Net Wins' score that ranks the cohort. The authors take the held-out results, together with the two ablations, as evidence that the intermediate structured report is doing real work: it supplies the second stage with faithful, comparable evidence and makes the reasoning auditable. The prediction is therefore cohort-relative rather than absolute, and the whole chain—images to text report to pairwise comparisons to ranking—is inspectable at every step.

Load-bearing premise

The load-bearing premise is that Stage 1's multimodal LLM, reading a montage of 128x128 node patches with overlaid labels, correctly and consistently identifies the ten radiological features; the paper does not validate these JSON reports against radiologist readings or node pathology, so if the reports misdescribe the images, the pairwise rankings and final scores inherit those errors.

Editorial extensions

If this is right

  • If the test-set result holds in larger cohorts, two-stage LLM reasoning becomes a viable alternative to end-to-end deep classifiers for lymph node staging.
  • The structured report makes predictions auditable: a clinician can see which nodes were flagged for which features and read the textual reason one patient was ranked higher than another.
  • The ablation results imply that neither stage is decoration: removing structured prompting or relational ranking drops AUC below most baselines, so the performance is attributable to the specific two-stage design.
  • Because Stage 2 operates only on text, the framework can in principle be adapted to any structured feature source, including radiologist-written reports, without retraining the visual stage.
  • The cohort-relative 'Net Wins' score may be better suited to triage and borderline-case flagging than a single absolute probability, since risk is defined against observed peers.

Reading between the lines

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

  • The paper does not separate perception error from reasoning error; a natural test is to compare Stage 1's JSON features with radiologist annotations on the same node patches, and if agreement is low while LRMR's AUC stays high, the performance would come from the comparison or montage structure rather than faithful feature extraction.
  • A second extension the paper does not run is measuring how stable the pairwise rankings are across random opponent draws and cohort composition; the 'Net Wins' score may shift with prevalence, which matters for deployment across centers.
  • The same two-stage pattern—structured multimodal report then pairwise text comparison—could transfer to other multi-lesion staging problems such as lung nodules or lateral pelvic nodes, where patient-level context is the deciding signal.
  • One could run Stage 2 on human radiologist reports in place of LLM reports to test the reasoning stage independently of visual perception, separating the framework's text-reasoning value from its image-to-text value.
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

5 major / 6 minor

Summary. The manuscript proposes LRMR, a two-stage LLM-based framework for patient-level lymph node metastasis assessment in rectal cancer. Stage 1 compiles all T2-weighted lymph node patches into a labeled montage and asks a multimodal LLM (Gemini 2.5 Flash) to output a structured JSON report on ten radiological features. Stage 2 feeds pairs of these reports to a text-only LLM for pairwise relative-risk comparisons, aggregates the outcomes into a per-patient 'net wins' score, and applies a threshold calibrated on the training set. On a retrospective cohort of 117 patients (81 train, 36 test), the authors report an AUC of 0.7917 and F1 of 0.7200, exceeding ResNet50 (AUC 0.7708), and ablations showing large drops when either the relational ranking stage or the structured prompt is removed.

Significance. If the claims were supported, the two-stage decoupling of visual perception from relational reasoning would be an interesting and clinically interpretable alternative to end-to-end black-box classifiers, and the structured-report design is a useful idea for auditing model decisions. The pipeline is clearly described, the ablation targets the two main contributions, and the paper is generally readable. However, the current evidence is insufficient to establish the headline claims: the evaluation rests on a single 36-patient test set without uncertainty quantification, the scoring procedure is stochastic and not reproducible as reported, and the visual feature extraction is unvalidated. The contribution is therefore better characterized as a promising but unvalidated pipeline than as a demonstrated improvement over existing methods.

major comments (5)
  1. [Section III-D, Table I] The central claim that LRMR outperforms the baselines is not statistically supported. The test set contains 36 patients (12 positive and 24 negative, as inferable from the reported accuracy, recall, and precision), and the AUC margin over ResNet50 is 0.0209 (0.7917 vs 0.7708). No confidence intervals, bootstrap estimates, or significance tests are reported, and with this sample size a margin of this size is well within sampling noise. The same problem applies to Table II: the ablation gaps of 0.1042 and 0.1459 AUC are single point estimates with no uncertainty, so the attribution of the gain to the two stages is not established.
  2. [Section II-C, Section III-B] The reported scores are one realization of a stochastic, unrepeatable procedure. For each patient, N=6 opponents are 'randomly sampled' without a stated seed or repeated draws, and the text-based LLM's temperature is not specified (only the Stage 1 multimodal LLM temperature of 0.1 is given). Consequently, the net-wins scores, the calibrated threshold, and all test-set metrics could change materially under a different random draw or LLM generation, and the reader cannot reproduce the numbers.
  3. [Section II-B] Stage 1's structured reports are the sole input to the relational ranking stage, yet the manuscript provides no validation that Gemini 2.5 Flash extracts the ten radiological features correctly from the 128x128-pixel montage patches. There is no comparison against radiologist labels, no node-level ground-truth check, and no analysis of hallucination or misattribution errors. If the reports misrepresent the images, every downstream ranking and the interpretability claim are compromised.
  4. [Section III-C, Section III-B] The comparison between LRMR and the deep learning baselines is confounded by model scale and pretraining. The baselines are trained from scratch on 81 patients using aggregated node features, whereas LRMR uses the pretrained Gemini 2.5 Flash model with substantial external knowledge. The observed gap may reflect pretraining and model capacity rather than the proposed two-stage design; an apples-to-apples comparison would require pretrained image encoders or otherwise matched capacity and training protocol.
  5. [Section II-C] The aggregation rule for the pairwise comparisons is underspecified. The model can output 'A', 'B', or 'Comparable', but the net-wins score is defined only as wins minus losses; the manuscript does not state how 'Comparable' decisions are counted, whether each unordered pair is judged once or twice, or how conflicting judgments of the same pair are reconciled.
minor comments (6)
  1. [Throughout] Figure and table references should be explicit: 'illustrated in 1' should be 'Fig. 1', '2a'/'2b' should be 'Fig. 2(a)'/'Fig. 2(b)', and 'presented in I' and 'in II' should be 'Table I' and 'Table II'.
  2. [Section II-B] The paper says the prompt contains 'ten specific questions' about radiological features but then lists only examples (shape, border, signal texture, fatty hilum, necrosis); the full list of ten features should be enumerated so the experiment is reproducible.
  3. [Section III-B] Please specify the Stage 2 LLM temperature, the number of repeated generations per comparison, any JSON parsing/retry logic, and the random seed or a multiple-seed protocol for opponent sampling.
  4. [References] Reference [3] (RIM-One, an open retinal image database) appears unrelated to lymph node staging in rectal cancer and should be replaced with an appropriate citation.
  5. [Section III-A] The dataset description should include ethics approval, inclusion/exclusion criteria, how lymph node patches were detected (manual annotation or automated method), and the range of node counts per patient; these details affect montage construction and generalizability.
  6. [Section III-E] In the 'LRMR w/o Structured Prompt' ablation, please clarify what the free-form Stage 1 output looked like and how it was fed into a Stage 2 prompt that expects structured JSON, so that the only changed component is indeed the prompt.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LRMR's evaluation chain is empirical and self-contained.

full rationale

The paper's central claim is an empirical performance comparison, not a mathematical derivation with equations that could reduce to inputs. The two-stage pipeline uses a multimodal LLM to convert montaged lymph-node images into structured JSON reports, then a text LLM performs pairwise report comparisons to produce a net-wins risk score. That score is evaluated against held-out histopathology labels via AUC and F1; no parameter of the model is fitted to the test set to produce the headline numbers. The only threshold calibration uses the training split, which is standard and does not make the test AUC circular. Self-citations in the reference list (e.g., Refs. [16]–[20], [25], [27]) are contextual related work and are not load-bearing premises for the framework's design or its reported superiority. There is no fitted input renamed as a prediction, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The methodological weaknesses highlighted by a skeptical reader—36-patient test set, stochastic pairwise sampling without seeds, and absence of confidence intervals—are threats to statistical reliability, not examples of circular reasoning. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities. Its central claim rests on unvalidated perceptual accuracy of the multimodal LLM, transitivity of pairwise comparisons, and the statistical adequacy of a 36-patient test set.

free parameters (4)
  • Number of comparison opponents per patient (N) = 6
    Chosen by hand in Section II-C; no sensitivity analysis. The random opponent set directly changes the net-wins score.
  • Classification threshold on net-wins score = not reported
    Selected on the 81-patient training set to produce accuracy/F1 (Section III-A); does not affect AUC but is a fitted quantity.
  • LLM temperature = 0.1
    Set for Stage 1 to encourage determinism; affects output variability; no repeat measurements across temperatures.
  • Montage patch size = 128x128 pixels
    Chosen for input standardization; may discard detail relevant to fine node features.
assumptions (5)
  • domain assumption The ten handcrafted radiological features are sufficient and correctly capture malignancy-relevant cues on T2-weighted MRI.
    Stage 1 prompts are limited to these features (Section II-B); no evidence they are complete or optimally discriminative.
  • domain assumption Gemini can perceive radiological details from a low-resolution montage with overlaid alphanumeric labels.
    Montage construction and label overlay (Section II-B) are unvalidated; LLM visual perception on such synthetic layouts is untested.
  • domain assumption LLM pairwise comparisons yield a transitive, consistent global ranking.
    Net-wins aggregation (Section II-C) assumes comparisons are coherent; no tests for intransitivity or inconsistency are reported.
  • domain assumption A 70/30 stratified split of 117 patients gives a reliable performance estimate.
    Test set has 36 patients; no confidence intervals or bootstrapping (Section III-A, Table I).
  • domain assumption The net-wins score is meaningful for a single new patient during deployment.
    The score is defined relative to a sampled opponent set (Section II-C); a standalone patient has no defined score.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LRMR: LLM-Driven Relational Multi-node Ranking for Lymph Node Metastasis Assessment in Rectal Cancer." pith.science (2026). https://pith.science/paper/C6F44A6C

@misc{pith2026250711457,
  author       = {Pith},
  title        = {Pith review of: LRMR: LLM-Driven Relational Multi-node Ranking for Lymph Node Metastasis Assessment in Rectal Cancer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6F44A6C}},
  note         = {Machine review of arXiv:2507.11457}
}
read the original abstract

Accurate preoperative assessment of lymph node (LN) metastasis in rectal cancer guides treatment decisions, yet conventional MRI evaluation based on morphological criteria shows limited diagnostic performance. While some artificial intelligence models have been developed, they often operate as black boxes, lacking the interpretability needed for clinical trust. Moreover, these models typically evaluate nodes in isolation, overlooking the patient-level context. To address these limitations, we introduce LRMR, an LLM-Driven Relational Multi-node Ranking framework. This approach reframes the diagnostic task from a direct classification problem into a structured reasoning and ranking process. The LRMR framework operates in two stages. First, a multimodal large language model (LLM) analyzes a composite montage image of all LNs from a patient, generating a structured report that details ten distinct radiological features. Second, a text-based LLM performs pairwise comparisons of these reports between different patients, establishing a relative risk ranking based on the severity and number of adverse features. We evaluated our method on a retrospective cohort of 117 rectal cancer patients. LRMR achieved an area under the curve (AUC) of 0.7917 and an F1-score of 0.7200, outperforming a range of deep learning baselines, including ResNet50 (AUC 0.7708). Ablation studies confirmed the value of our two main contributions: removing the relational ranking stage or the structured prompting stage led to a significant performance drop, with AUCs falling to 0.6875 and 0.6458, respectively. Our work demonstrates that decoupling visual perception from cognitive reasoning through a two-stage LLM framework offers a powerful, interpretable, and effective new paradigm for assessing lymph node metastasis in rectal cancer.

Figures

Figures reproduced from arXiv: 2507.11457 by the authors.

Figure 1
Figure 1. The overall pipeline of the proposed framework. For each patient, all T2-weighted lymph node images are first compiled into a single montage image. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Detailed illustration of the two-stage LRMR framework. (a) Stage 1: Structured Feature Extraction. All lymph node patches from a patient are compiled [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 26 canonical work pages

  1. [1]

    Lymph nodes evaluation in rectal cancer: where do we stand and future perspective,

    A. Borgheresi, F. De Muzio, A. Agostini, L. Ottaviani, A. Bruno, V . Granata, R. Fusco, G. Danti, F. Flammia, R. Grassi et al. , “Lymph nodes evaluation in rectal cancer: where do we stand and future perspective,” Journal of Clinical Medicine, vol. 11, no. 9, p. 2599, 2022

    work page 2022

  2. [2]

    Assessment of lymph node involvement in colorectal cancer,

    M. L. Ong and J. B. Schofield, “Assessment of lymph node involvement in colorectal cancer,” World journal of gastrointestinal surgery , vol. 8, no. 3, p. 179, 2016

    work page 2016

  3. [3]

    Rim-one: An open retinal image database for optic nerve evaluation,

    F. Fumero, S. Alay ´on, J. L. Sanchez, J. Sigut, and M. Gonzalez- Hernandez, “Rim-one: An open retinal image database for optic nerve evaluation,” in 2011 24th international symposium on computer-based medical systems (CBMS) . IEEE, 2011, pp. 1–6

    2011

  4. [4]

    Lymph node metastasis in cancer progression: molecular mechanisms, clinical significance and therapeutic interventions,

    H. Ji, C. Hu, X. Yang, Y . Liu, G. Ji, S. Ge, X. Wang, and M. Wang, “Lymph node metastasis in cancer progression: molecular mechanisms, clinical significance and therapeutic interventions,” Signal Transduction and Targeted Therapy, vol. 8, no. 1, p. 367, 2023

    work page 2023

  5. [5]

    Prognostic value of clinical vs pathologic stage in rectal cancer patients receiving neoadjuvant therapy,

    D. Delitto, T. J. George Jr, T. J. Loftus, P. Qiu, G. J. Chang, C. J. Allegra, W. A. Hall, S. J. Hughes, S. A. Tan, C. M. Shaw et al. , “Prognostic value of clinical vs pathologic stage in rectal cancer patients receiving neoadjuvant therapy,” JNCI: Journal of the National Cancer Institute , vol. 110, no. 5, pp. 460–466, 2018

    work page 2018

  6. [6]

    Imaging and metabolic diagnostic methods in the stage assessment of rectal cancer,

    R. Maksim, A. Buczy ´nska, I. Sidorkiewicz, A. J. Kretowski, and E. Sierko, “Imaging and metabolic diagnostic methods in the stage assessment of rectal cancer,” Cancers, vol. 16, no. 14, p. 2553, 2024

    work page 2024

  7. [7]

    Deep learning models for preoperative t-stage assessment in rectal cancer using mri: exploring the impact of rectal filling,

    C. Tian, X. Ma, H. Lu, Q. Wang, C. Shao, Y . Yuan, and F. Shen, “Deep learning models for preoperative t-stage assessment in rectal cancer using mri: exploring the impact of rectal filling,” Frontiers in Medicine, vol. 10, p. 1326324, 2023

    work page 2023

  8. [8]

    Comparing abbreviated and full mri protocols for preoperative local staging of locally advanced rectal cancer,

    N. Y . Ebaid, S. E. Badr, R. F. Mansour, H. A. Abo-Alella, M. M. Assy, S. K. S. Eldemerdash, M. A. S. A. Haasan, and H. A. E. Mohamed, “Comparing abbreviated and full mri protocols for preoperative local staging of locally advanced rectal cancer,” Academic Radiology, 2025

    work page 2025

Show all 36 references
  1. [9]

    Mri of rectal cancer: tumor staging, imaging techniques, and management,

    N. Horvat, C. Carlos Tavares Rocha, B. Clemente Oliveira, I. Petkovska, and M. J. Gollub, “Mri of rectal cancer: tumor staging, imaging techniques, and management,” Radiographics, vol. 39, no. 2, pp. 367– 387, 2019

    2019

  2. [10]

    Local staging of rectal cancer: a review of imaging,

    R. G. Beets-Tan and G. L. Beets, “Local staging of rectal cancer: a review of imaging,” Journal of Magnetic Resonance Imaging , vol. 33, no. 5, pp. 1012–1019, 2011

    2011

  3. [11]

    Magnetic resonance imaging evaluation of the accuracy of various lymph node staging criteria in rectal cancer: a systematic review and meta-analysis,

    Z. Zhuang, Y . Zhang, M. Wei, X. Yang, and Z. Wang, “Magnetic resonance imaging evaluation of the accuracy of various lymph node staging criteria in rectal cancer: a systematic review and meta-analysis,” Frontiers in Oncology, vol. 11, p. 709070, 2021

    2021

  4. [12]

    Micrometastasis in lymph nodes of colorectal cancer,

    H. Yamamoto, “Micrometastasis in lymph nodes of colorectal cancer,” Annals of Gastroenterological Surgery, vol. 6, no. 4, pp. 466–473, 2022

    2022

  5. [13]

    The multimodal management of locally advanced rectal can- cer: making sense of the new data,

    W. H. Zwart, A. Hotca, G. A. Hospers, K. A. Goodman, and J. Garcia- Aguilar, “The multimodal management of locally advanced rectal can- cer: making sense of the new data,” American Society of Clinical Oncology Educational Book , vol. 42, pp. 1–14, 2022

    2022

  6. [14]

    The role of preoperative imaging in the detection of lateral lymph node metastases in rectal cancer: a systematic review and diagnostic test meta-analysis,

    S. Rooney, J. Meyer, Z. Afzal, J. Ashcroft, H. Cheow, K. DePaepe, M. Powar, C. Simillis, J. Wheeler, J. Davies et al. , “The role of preoperative imaging in the detection of lateral lymph node metastases in rectal cancer: a systematic review and diagnostic test meta-analysis,”...

    2022

  7. [15]

    Diag- nostic accuracy of lateral lymph node metastasis for locally advanced rectal cancer after neoadjuvant therapy: a systematic review and meta- analysis,

    K. Ozaki, K. Kawai, S. Ogawa, Y . Kanemitsu, and Y . Ajioka, “Diag- nostic accuracy of lateral lymph node metastasis for locally advanced rectal cancer after neoadjuvant therapy: a systematic review and meta- analysis,” Expert Review of Anticancer Therapy , pp. 1–7, 2025

    2025

  8. [16]

    An anatomy-aware frame- work for automatic segmentation of parotid tumor from multimodal mri,

    Y . Gao, Y . Dai, F. Liu, W. Chen, and L. Shi, “An anatomy-aware frame- work for automatic segmentation of parotid tumor from multimodal mri,” Computers in Biology and Medicine , vol. 161, p. 107000, 2023

    2023

  9. [17]

    Prior-driven refinement network for small organ-at-risk segmentation in head and neck cancer,

    T. Wang, Y . Gao, B. Liang, and Q. Wang, “Prior-driven refinement network for small organ-at-risk segmentation in head and neck cancer,” Engineering Applications of Artificial Intelligence , vol. 159, p. 111605, 2025

    2025

  10. [18]

    Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation,

    Y . Gao, W. Xia, W. Wang, and X. Gao, “Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation,” in Interna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 437–447

    2024

  11. [19]

    Transmed: Transformers advance multi- modal medical image classification,

    Y . Dai, Y . Gao, and F. Liu, “Transmed: Transformers advance multi- modal medical image classification,” Diagnostics, vol. 11, no. 8, p. 1384, 2021

    2021

  12. [20]

    Safeclick: Error-tolerant interactive segmentation of any medical volumes via hierarchical expert consensus,

    Y . Gao, J. Sheng, W. Wu, H. Li, Y . Dong, C. Ge, F. Yuan, and X. Gao, “Safeclick: Error-tolerant interactive segmentation of any medical volumes via hierarchical expert consensus,” arXiv preprint arXiv:2506.18404, 2025

    2025 arXiv

  13. [21]

    An unsupervised anatomy-aware dual-constraint cascade network for lung computed tomography deformable image registration,

    W. Wu, Y . Gao, X. Jin, R. Zhang, Y . Pan, and X. Gao, “An unsupervised anatomy-aware dual-constraint cascade network for lung computed tomography deformable image registration,” Engineering Applications of Artificial Intelligence , vol. 158, p. 111548, 2025

    2025

  14. [22]

    Desam: Decoupled seg- ment anything model for generalizable medical image segmentation,

    Y . Gao, W. Xia, D. Hu, W. Wang, and X. Gao, “Desam: Decoupled seg- ment anything model for generalizable medical image segmentation,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 509–519

    2024

  15. [23]

    Tooth instance segmentation and disease detection with uncertainty-aware contrastive learning and cross-scale attention,

    X. Xu, J. Chen, and J. Yin, “Tooth instance segmentation and disease detection with uncertainty-aware contrastive learning and cross-scale attention,” IEEE Journal of Biomedical and Health Informatics , 2025

    2025

  16. [24]

    Emcnet: Ensemble multiscale convolutional neural network for single-lead ecg classification in wearable devices,

    J. Chen, B. Fang, H. Li, L.-B. Zhang, Y . Teng, and G. Fortino, “Emcnet: Ensemble multiscale convolutional neural network for single-lead ecg classification in wearable devices,” IEEE Sensors Journal, vol. 24, no. 6, pp. 8754–8762, 2024

    2024

  17. [25]

    Multimodal large language models address clinical queries in laryngeal cancer surgery: a comparative evaluation of image interpretation across different models,

    B. Liang, Y . Gao, T. Wang, L. Zhang, and Q. Wang, “Multimodal large language models address clinical queries in laryngeal cancer surgery: a comparative evaluation of image interpretation across different models,” International Journal of Surgery , vol. 111, no. 3, pp. 2727–2730, 2025

    2025

  18. [26]

    Multicenter evaluation of a weakly supervised deep learning model for lymph node diagnosis in rectal cancer at mri,

    W. Xia, D. Li, W. He, P. J. Pickhardt, J. Jian, R. Zhang, J. Zhang, R. Song, T. Tong, X. Yang et al., “Multicenter evaluation of a weakly supervised deep learning model for lymph node diagnosis in rectal cancer at mri,” Radiology: Artificial Intelligence , vol. 6, no. 2, p. e2...

    2024

  19. [27]

    Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer,

    Y . Gao, Y . Dong, W. Wu, C. Ge, F. Yuan, J. Sheng, H. Li, and X. Gao, “Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer,” arXiv preprint arXiv:2505.10502, 2025

    2025 arXiv

  20. [28]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in Proceedings of the IEEE confer- ence on computer vision and pattern recognition , 2017, pp. 4700–4708

    2017

  21. [29]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” in International conference on machine learning. PMLR, 2019, pp. 6105–6114

    2019

  22. [30]

    Deep residual learning for image recognition,

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

    2016

  23. [31]

    A convnet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 11 976–11 986

    2022

  24. [32]

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

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

    2010 arXiv

  25. [33]

    Beit: Bert pre-training of image transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “Beit: Bert pre-training of image transformers,” arXiv preprint arXiv:2106.08254 , 2021

    2021 arXiv

  26. [34]

    Mlp-mixer: An all-mlp architecture for vision,

    I. O. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Un- terthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreitet al., “Mlp-mixer: An all-mlp architecture for vision,” Advances in neural information processing systems, vol. 34, pp. 24 261–24 272, 2021

    2021

  27. [35]

    Resmlp: Feed- forward networks for image classification with data-efficient training,

    H. Touvron, P. Bojanowski, M. Caron, M. Cord, A. El-Nouby, E. Grave, G. Izacard, A. Joulin, G. Synnaeve, J. Verbeek et al. , “Resmlp: Feed- forward networks for image classification with data-efficient training,” IEEE transactions on pattern analysis and machine intelligence, ...

    2022

  28. [36]

    Mambaout: Do we really need mamba for vision?

    W. Yu and X. Wang, “Mambaout: Do we really need mamba for vision?” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4484–4496

    2025

Pith tools

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