Pith. sign in

REVIEW 3 major objections 33 references

Conditioning an LLM on high-level semantic differences between a CT scan and a normal reference yields more accurate, clinically detailed 3D CT reports without explicit lesion localization.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 23:00 UTC pith:XMSP6GVR

load-bearing objection Solid 3D CT report-gen method paper: hierarchical scan-to-normal latent differences as LLM visual prefixes work, but clinical gains need clean in-domain normals. the 3 major comments →

arxiv 2603.17717 v4 pith:XMSP6GVR submitted 2026-03-18 cs.CR cs.AIstat.APstat.ML

Machine Learning for Network Attacks Classification and Statistical Evaluation of Adversarial Learning Methodologies for Synthetic Data Generation

Iakovos-Christos Zarkadis , Christos Douligeris This is my paper
classification cs.CR cs.AIstat.APstat.ML
keywords CT report generationvision-language modelssemantic discrepancy modelingvisual promptinglarge language models3D medical imagingclinical efficacy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Radiologists write CT reports by comparing a scan to an internal sense of normal anatomy, not by treating every voxel as equally informative. Most LLM-based report generators ignore that habit: they dump the whole volume as visual tokens and hope the model finds what matters. This paper argues that you can close the gap by extracting hierarchical global and local semantic differences between the target scan and a normal reference, then turning those differences into short learnable visual prefix tokens that condition the LLM. On two large chest-CT benchmarks the approach improves language metrics and clinical-efficacy scores over strong prior methods, while remaining robust to which normal reference is paired at random. The practical payoff is reports that better surface sparse findings without requiring lesion segmentation or pixel-perfect registration.

Core claim

Explicit scan-to-normal semantic discrepancy, modeled at complementary global and local scales in a shared latent space and injected as learnable visual prefix tokens, steers an LLM toward diagnostically relevant evidence and produces better CT reports than feeding absolute volume tokens alone.

What carries the argument

Differential Visual Prompting (DiffVP): a hierarchical difference extractor that computes a global delta via a Transformer query over resampled target and reference tokens plus a local delta via distance-weighted residual aggregation, then a difference-to-prompt generator that projects the fused deltas into a short sequence of soft prefix tokens for the LLM.

Load-bearing premise

A randomly paired normal CT from a curated “no abnormality” pool must be clean and domain-matched enough that latent differences reflect disease rather than scanner, protocol, or anatomy noise.

What would settle it

Repeat the RadGenome-ChestCT evaluation with the same architecture but only cross-domain or deliberately contaminated normal pools; if language and clinical-efficacy gains over the no-difference baseline vanish, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • LLM-based CT report systems can gain sensitivity to sparse findings without adding lesion detectors or region masks.
  • High-level semantic comparison is preferable to pixel-level subtraction when inter-subject anatomy and registration error make raw difference maps noisy.
  • A modest in-domain normal reference pool, even if small, is enough to support the contrast mechanism in deployment.
  • Prefix-style difference prompts offer a controllable way to reweight visual conditioning so invariant anatomy is downplayed and deviant regions are amplified.

Where Pith is reading between the lines

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

  • The same hierarchical difference-prefix pattern could transfer to other sparse 3D imaging tasks (MRI, PET) where a normative atlas or healthy prior exists.
  • If clean normals are scarce, synthesizing or adapting normal priors may become as important as improving the generator itself.
  • Ordering-sensitive metrics such as ROUGE-L may systematically undervalue clinically correct reports that rearrange findings, so clinical-efficacy labels remain the sharper test of this approach.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper proposes DiffVP, a reference-guided framework for LLM-based 3D CT report generation. Instead of feeding whole-volume visual tokens uniformly into an LLM, DiffVP contrasts a target scan with a normal reference in a shared latent space via a Hierarchical Difference Extractor (global Transformer difference query plus distance-weighted local residuals) and maps the fused discrepancy into learnable visual prefix tokens through a Difference-to-Prompt Generator. These prefixes, optionally combined with a frozen 18-class diagnostic anchor, condition a LoRA-tuned LLaMA-2-7B. On RadGenome-ChestCT and CTRG-Chest-548K the method reports large gains over prior specialized and foundation models (e.g., BLEU-1/4 and clinical-efficacy F1 0.421 on RadGenome), supported by module ablations, prefix-length sweeps, semantic-vs-pixel difference comparisons, five-seed random-reference stability with t-tests, and sensitivity analyses of normal-pool size, domain shift, and contamination.

Significance. If the results hold under the stated boundary conditions, DiffVP is a clear and practically relevant contribution: it operationalizes radiologists’ comparative reasoning without explicit lesion localization, and it shows that high-level semantic discrepancy prompts can suppress anatomical redundancy and improve both n-gram metrics and clinical recall. Strengths include a coherent two-module design, multi-metric evaluation (NLG + RadBERT CE), honest documentation of normal-pool sensitivity (Appendix F), multi-seed significance testing (Table 5), and a commitment to release code. The work is therefore of interest to medical VLM and report-generation communities, provided the dependence on clean in-domain normals and any internal ablation inconsistencies are resolved or clearly scoped.

major comments (3)
  1. Table 4 vs Table 6 (Appendix D) report incompatible clinical-efficacy numbers for the same module setting. Table 4 lists global+local (without E, if the checkmarks are read as written) with CE-F1 ≈ 0.411 (Rec 0.459), whereas Table 6 reports +Δglobal+Δlocal (w/o E) with CE-F1 = 0.238 (Rec 0.175). The narrative in §5.1 also does not uniquely determine which of #2/#3 include E. This is load-bearing for the claim that hierarchical difference modeling—not the diagnostic anchor—drives clinical gains. Please reconcile the tables, state the exact module configuration for every row, and re-run or re-export the ablations so that main-text and appendix numbers agree.
  2. §4.1 and Appendix F establish that clinical efficacy collapses under cross-domain or contaminated normal pools (CE-F1 ~0.42 → ~0.13; Tables 8–10), while NLG remains only modestly above the no-Δ baseline. The central claim that DiffVP amplifies diagnostically meaningful evidence therefore rests on access to a clean, in-domain “no abnormality” reference pool. This boundary condition is currently concentrated in the appendix; it should be stated explicitly in the abstract/introduction/limitations and, if possible, stress-tested with a practical protocol (e.g., few-shot in-domain normals only, or a learned normal prototype) so readers can judge deployability.
  3. Eqs. (3)–(7) and §3.3 assume that shared Q-Former queries project unregistered, inter-subject volumes into a token space where Euclidean residuals and a global difference query are clinically meaningful. The paper shows robustness to random pairing within a clean pool (Table 5) but does not quantify residual anatomy/scanner mismatch when the reference is only “report-normal.” A controlled analysis (e.g., same-patient prior vs random normal, or registration-error injection) would strengthen the claim that the hierarchical operators isolate pathology rather than protocol/anatomy noise—especially given the domain-shift collapse already reported.

Circularity Check

0 steps flagged

No significant circularity: DiffVP is an empirical architecture paper whose gains are measured on held-out data with external metrics, not forced by definition or self-citation.

full rationale

The manuscript proposes Hierarchical Difference Extractor + Difference-to-Prompt Generator modules that map scan-to-normal latent discrepancies into LLM prefix tokens, then reports BLEU/METEOR/ROUGE and RadBERT-based clinical-efficacy scores on held-out RadGenome-ChestCT and CTRG-Chest-548K splits. Normal references are drawn only from the matching split (train-only / test-only) to avoid leakage; the 18-class diagnostic anchor is a frozen ResNet-18 trained on structured labels, not on free-text reports; CE evaluation uses a separate RadBERT pipeline. Ablations (global/local delta, prefix length, no-Δ / pixel-Δ / latent-Δ) and multi-seed stability tests treat the method as an empirical hypothesis, not a quantity derived from its own evaluation targets. Related-work citations motivate design choices but do not supply a uniqueness theorem or load-bearing self-citation that forces the reported gains. The documented sensitivity of CE-F1 to cross-domain or contaminated normal pools is a boundary condition on the method’s assumption, not a circular reduction of prediction to input. No equation or claim reduces by construction to a fitted parameter renamed as a first-principles result.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim rests on standard VLM engineering choices plus domain assumptions about normal references and evaluation proxies. Free parameters are architectural/training knobs selected on validation performance (prefix length 16, latent token budget, LoRA, etc.). Invented entities are modules (HDE, DPG, difference prompts), not physical objects; their evidence is internal ablations rather than external independent measurement.

free parameters (5)
  • visual prefix length p
    Chosen by sweep over {4,8,16,24,32}; peak at p=16 on RadGenome NLG metrics and used as default for main claims.
  • resampler latent token count N / shared queries
    Fixed-budget latent size for Q-Former-style resampling; controls alignment granularity of target vs reference comparison.
  • learning rate 5e-5, 10 epochs, batch size 1, LoRA ranks (unspecified detail)
    Optimization hyperparameters that affect final BLEU/CE; standard but claim-dependent fitted choices.
  • local residual distance epsilon (1e-5)
    Numerical stabilizer in importance-weight normalization for local delta; minor but part of the local operator definition.
  • multi-task loss combination L_gen + L_cls (implicit weight 1:1)
    Equal sum of generation NLL and auxiliary 18-class BCE; weighting is a design choice affecting CE vs NLG tradeoff.
axioms (5)
  • domain assumption High-level latent differences between a target CT and a normal reference are more clinically informative for report generation than absolute single-scan tokens or pixel-level subtraction.
    Core modeling premise stated in Introduction/Methodology; supported by ablations but not independently proven as a general radiology principle.
  • domain assumption Reports labeled as no-abnormality define a usable normal reference pool for random pairing without train-test leakage.
    §4.1 and Appendix B construct pools from report text; Appendix F shows sensitivity when this cleanliness/domain match fails.
  • ad hoc to paper Shared learnable queries + cross-attention project different subjects into a comparable latent token space despite inter-subject anatomy and lack of registration.
    Enables token-wise local residuals and global difference query without spatial registration (Eqs. 3–7).
  • domain assumption NLG metrics plus RadBERT-extracted 18-label clinical efficacy are adequate proxies for report quality and diagnostic correctness.
    Standard in the subfield; paper itself notes NLG insufficiency and CE label-alignment limits on CTRG.
  • standard math Transformer/attention, LoRA fine-tuning, and prefix-tuning are valid conditioning mechanisms for LLaMA-2 report decoding.
    Background ML machinery cited via LoRA and prefix-tuning literature.
invented entities (3)
  • Hierarchical Difference Extractor (global delta query + local distance-weighted residual) no independent evidence
    purpose: Compute multi-scale semantic discrepancies between target and normal CT in latent token space.
    New module composition for this task; evidence is ablation performance, not external theory.
  • Difference-to-Prompt Generator / visual difference prefix tokens no independent evidence
    purpose: Map fused global/local deltas into LLM embedding-space soft prefixes that reweight conditioning.
    Explicit prefix injection of 3D discrepancy is the paper’s distinctive interface to the LLM.
  • Diagnostic anchor E (frozen 18-class 3D ResNet-18 disease prompt) independent evidence
    purpose: Provide high-level disease text anchors without free-text leakage.
    Authors note this is a standard paradigm, not core novelty; still load-bearing for best CE.

pith-pipeline@v1.1.0-grok45 · 23174 in / 3902 out tokens · 46289 ms · 2026-07-13T23:00:20.016213+00:00 · methodology

0 comments
read the original abstract

Supervised detection of network attacks has always been a critical part of network intrusion detection systems (NIDS). Nowadays, in a pivotal time for artificial intelligence (AI), with even more sophisticated attacks that utilize advanced techniques, such as generative artificial intelligence (GenAI) and reinforcement learning, it has become a vital component if we wish to protect our personal data, which are scattered across the web. In this paper, we address two tasks, in the first unified multi-modal NIDS dataset, which incorporates flow-level data, packet payload information and temporal contextual features, from the reprocessed CIC-IDS-2017, CIC-IoT-2023, UNSW-NB15 and CIC-DDoS-2019, with the same feature space. In the first task we use machine learning (ML) algorithms, with stratified cross validation, in order to prevent network attacks, with stability and reliability. In the second task we use adversarial learning algorithms to generate synthetic data, compare them with the real ones and evaluate their fidelity, utility and privacy using the SDV framework, f-divergences, distinguishability and non-parametric statistical tests. The findings provide stable ML models for intrusion detection and generative models with high fidelity and utility, by combining the Synthetic Data Vault framework, the TRTS and TSTR tests, with non-parametric statistical tests and f-divergence measures.

Figures

Figures reproduced from arXiv: 2603.17717 by Christos Douligeris, Iakovos-Christos Zarkadis.

Figure 1
Figure 1. Figure 1: XGBoost Precision-Recall Stratified 10-Fold Cross Validation. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: GAN and WGAN Loss Progress. Furthermore we highlight the loss progress of TVAE, whose results are the highest, compared to the rest non-conditional architectures. Although it fails generating equal number of observations per network attack. It’s performance could be further optimized with the use of a conditional architecture and a mixture approach [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: TVAE Results C. Evaluation For evaluating the quality of the new synthetically generated data we chose the Synthetic Data Vault framework (SDV) [26]. From the quality report, we investigated the statistical fidelity, measuring how well we have captured the marginal distribution of each feature using the Kolmogorov-Smirnov Complement and how well the correlations have been pre￾served using the Correlation S… view at source ↗
Figure 4
Figure 4. Figure 4: We also performed statistical hypothesis tests in order to compare the multivariate means and the covariance matrices [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: between the real and the synthetic data. We utilized the Regularized Hotelling’s T 2 test [19] and the Frobenius Norm Covariance test [14], [16], for the multivariate means and covariance test respectively, each of which was computed for a total of 500 permutations, to conclude to the p − values seen in Tab. X. We consider a significance level of a = 5%, for both statistical tests seen in “(8)” and “(9)”. … view at source ↗

Review history (2 revisions) →

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

33 extracted references · 9 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    arXiv preprint arXiv:2404.00578 (2024) 16 Y

    Bai, F., Du, Y., Huang, T., Meng, M.Q.H., Zhao, B.: M3d: Advancing 3d medical image analysis with multi-modal large language models. arXiv preprint arXiv:2404.00578 (2024) 16 Y. Tian et al

    Pith/arXiv arXiv 2024

  3. [3]

    In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization

    Banerjee, S., Lavie, A.: Meteor: An automatic metric for mt evaluation with im- proved correlation with human judgments. In: Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summa- rization. pp. 65–72 (2005)

    2005

  4. [4]

    arXiv preprint arXiv:2010.16056 (2020)

    Chen, Z., Song, Y., Chang, T.H., Wan, X.: Generating radiology reports via memory-driven transformer. arXiv preprint arXiv:2010.16056 (2020)

    Pith/arXiv arXiv 2010

  5. [5]

    IEEE Transactions on Medical Imaging (2025)

    Chen, Z., Bie, Y., Jin, H., Chen, H.: Large language model with region-guided referring and grounding for ct report generation. IEEE Transactions on Medical Imaging (2025)

    2025

  6. [6]

    In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention

    Chen, Z., Luo, L., Bie, Y., Chen, H.: Dia-llama: Towards large language model- driven ct report generation. In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention. pp. 141–151. Springer (2025)

    2025

  7. [7]

    IEEE Journal of Biomedical and Health Informatics (2025)

    Deng, X., He, X., Bao, J., Zhou, Y., Cai, S., Cai, C., Chen, Z.: Mvketr: chest ct report generation with multi-view perception and knowledge enhancement. IEEE Journal of Biomedical and Health Informatics (2025)

    2025

  8. [8]

    CoRR (2024)

    Hamamci, I.E., Er, S., Almas, F., Simsek, A.G., Esirgun, S.N., Dogan, I., Dasdelen, M.F., Wittmann, B., Simsar, E., Simsar, M., et al.: A foundation model utilizing chest ct volumes and radiology reports for supervised-level zero-shot detection of abnormalities. CoRR (2024)

    2024

  9. [9]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Hamamci, I.E., Er, S., Menze, B.: Ct2rep: Automated radiology report generation for 3d medical imaging. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 476–486. Springer (2024)

    2024

  10. [10]

    ICLR1(2), 3 (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR1(2), 3 (2022)

    2022

  11. [11]

    In: Findings of the Association for Computational Linguistics: ACL 2023

    Kale, K., Bhattacharyya, P., Jadhav, K.: Replace and report: Nlp assisted radiology report generation. In: Findings of the Association for Computational Linguistics: ACL 2023. pp. 10731–10742 (2023)

    2023

  12. [12]

    Radiology116(3), 527–532 (1975)

    Kundel, H.L., Nodine, C.F.: Interpreting chest radiographs without visual search. Radiology116(3), 527–532 (1975)

    1975

  13. [13]

    Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Pro- cessing (Volume 1: Long Papers). pp. 4582–4597 (2021)

    2021

  14. [14]

    In: Text sum- marization branches out

    Lin, C.Y.: Rouge: A package for automatic evaluation of summaries. In: Text sum- marization branches out. pp. 74–81 (2004)

    2004

  15. [15]

    In: Findings of the Association for Computational Linguistics: ACL 2025

    Liu, C., Wan, Z., Wang, Y., Shen, H., Wang, H., Zheng, K., Zhang, M., Arcucci, R.: Argus: benchmarking and enhancing vision-language models for 3d radiology report generation. In: Findings of the Association for Computational Linguistics: ACL 2025. pp. 16448–16460 (2025)

    2025

  16. [16]

    Knowledge-Based Systems305, 112630 (2024)

    Lyu, C., Qiu, C., Han, K., Li, S., Sheng, V.S., Rong, H., Song, Y., Liu, Y., Liu, Z.: Automatic medical report generation combining contrastive learning and feature difference. Knowledge-Based Systems305, 112630 (2024)

    2024

  17. [17]

    arXiv preprint arXiv:2505.16229 (2025)

    Mao, Y., Xu, W., Qin, Y., Gao, Y.: Ct-agent: A multimodal-llm agent for 3d ct radiology question answering. arXiv preprint arXiv:2505.16229 (2025)

    Pith/arXiv arXiv 2025

  18. [18]

    Journal of Imag- ing Informatics in Medicine37(5), 2669–2687 (2024)

    Marcos, L., Babyn, P., Alirezaie, J.: Pure vision transformer (ct-vit) with noise2neighbors interpolation for low-dose ct image denoising. Journal of Imag- ing Informatics in Medicine37(5), 2669–2687 (2024)

    2024

  19. [19]

    In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311–318 (2002) DiffVP for LLM-Based CT Report Generation 17

    2002

  20. [20]

    In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop

    Park, H., Kim, K., Yoon, J., Park, S., Choi, J.: Feature difference makes sense: a medical image captioning model exploiting feature difference and tag information. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop. pp. 95–102 (2020)

    2020

  21. [21]

    arXiv preprint arXiv:2505.03401 (2025)

    Song, S., Tang, H., Yang, H., Li, X.: Ddatr: Dynamic difference-aware tempo- ral residual network for longitudinal radiology report generation. arXiv preprint arXiv:2505.03401 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Expert Systems with Applications237, 121442 (2024)

    Tang, Y., Yang, H., Zhang, L., Yuan, Y.: Work like a doctor: Unifying scan localizer and dynamic generator for automated computed tomography report generation. Expert Systems with Applications237, 121442 (2024)

    2024

  23. [23]

    arXiv preprint arXiv:2307.09288 (2023)

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bash- lykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    Pith/arXiv arXiv 2023

  24. [24]

    Meta-Radiology1(3), 100033 (2023)

    Wang, Z., Liu, L., Wang, L., Zhou, L.: R2gengpt: Radiology report generation with frozen llms. Meta-Radiology1(3), 100033 (2023)

    2023

  25. [25]

    In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

    Wang, Z., Tang, M., Wang, L., Li, X., Zhou, L.: A medical semantic-assisted trans- former for radiographic report generation. In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention. pp. 655–664. Springer (2022)

    2022

  26. [26]

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

    Wang, Z., Zhou, L., Wang, L., Li, X.: A self-boosting framework for automated radiographic report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2433–2442 (2021)

    2021

  27. [27]

    Nature Communications16(1), 7866 (2025)

    Wu, C., Zhang, X., Zhang, Y., Hui, H., Wang, Y., Xie, W.: Towards generalist foun- dation model for radiology by leveraging web-scale 2d&3d medical data. Nature Communications16(1), 7866 (2025)

    2025

  28. [28]

    Radiology: Artificial Intelligence4(4), e210258 (2022)

    Yan, A., McAuley, J., Lu, X., Du, J., Chang, E.Y., Gentili, A., Hsu, C.N.: Radbert: adapting transformer-based language models to radiology. Radiology: Artificial Intelligence4(4), e210258 (2022)

    2022

  29. [29]

    arXiv preprint arXiv:2407.10670 (2024)

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., et al.: Qwen2 technical report. arXiv preprint arXiv:2407.10670 (2024)

    Pith/arXiv arXiv 2024

  30. [30]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Yun, H., Maeng, J., Kang, E., Suk, H.I.: Diff-rrg: Longitudinal disease-wise patch difference as guidance for llm-based radiology report generation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 152–161. Springer (2025)

    2025

  31. [31]

    arXiv preprint arXiv:2404.16754 (2024)

    Zhang, X., Wu, C., Zhao, Z., Lei, J., Zhang, Y., Wang, Y., Xie, W.: Radgenome- chest ct: A grounded vision-language dataset for chest ct analysis. arXiv preprint arXiv:2404.16754 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    In: Findings of the Association for Computational Linguistics: EMNLP 2023

    Zhao, G., Yan, Y., Zhao, Z.: Normal-abnormal decoupling memory for medical report generation. In: Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 1962–1977 (2023)

    2023

  33. [33]

    arXiv preprint arXiv:2312.17183 (2023) 18 Y

    Zhao, Z., Zhang, Y., Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: One model to rule them all: Towards universal segmentation for medical images with text prompts. arXiv preprint arXiv:2312.17183 (2023) 18 Y. Tian et al. A Experiments Background ��� �������� ������������������Derived from the large-scale CT-RATE [8] dataset, which is specifically const...

    Pith/arXiv arXiv 2023