LoGSAM: Parameter-Efficient Cross-Modal Grounding for MRI Segmentation

Andreas Maier; Melika Qahqaie; Mohammad Robaitul Islam Bhuiyan; Paula Andrea P\'erez-Toro; Sheethal Bhat; Tom\'as Arias-Vergara; Tri-Thien Nguyen

REVIEW 4 major objections 6 minor 21 references

Radiologist dictation drives brain-tumor segmentation through a modular chain of pretrained speech, language, and vision models, reaching 98.6% of full fine-tuning while training under 5% of parameters.

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 · deepseek-v4-flash

2026-08-02 17:55 UTC pith:ZSYWDKJV

load-bearing objection A sensible modular speech-to-segmentation pipeline with real parameter-efficiency numbers, but the headline Dice is measured with clean text prompts, not with the speech-derived prompts the system is named for. the 4 major comments →

arxiv 2603.17576 v4 pith:ZSYWDKJV submitted 2026-03-18 cs.CV

LoGSAM: Parameter-Efficient Cross-Modal Grounding for MRI Segmentation

Mohammad Robaitul Islam Bhuiyan , Sheethal Bhat , Melika Qahqaie , Andreas Maier , Tri-Thien Nguyen , Paula Andrea P\'erez-Toro , Tom\'as Arias-Vergara This is my paper

classification cs.CV

keywords MRI segmentationbrain tumorspeech-to-segmentationparameter-efficient fine-tuningLoRAGrounding DINOMedSAMnegation-aware NLP

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.

LoGSAM tries to show that a speech-to-segmentation pipeline for brain tumors can be assembled from pretrained foundation models with almost no task-specific training. It transcribes radiologist dictation, extracts tumor-class prompts with negation handling, localizes tumors with a LoRA-adapted grounding detector, and lets a frozen medical segmenter produce masks from the resulting boxes. The central numbers are a mean Dice of 0.8032 on BRISC 2025 (98.6% of a fully fine-tuned baseline) while training under 5% of parameters, and correct tumor-class extraction from 11 of 12 German dictations. A sympathetic reader would care because, if true, it lowers the annotation and compute burden for medical segmentation and opens a route from routine clinical dictation to automatic masks.

Core claim

The paper's central discovery is that a fully modular chain — speech transcription and translation, negation-aware clinical NLP, a LoRA-tuned Grounding DINO detector for text-conditioned box prediction, and a frozen MedSAM for mask generation — can turn radiologist dictation into brain-tumor segmentations while adapting fewer than 5% of the parameters of the localization model. On BRISC 2025 it reaches a mean Dice of 0.8032, 98.6% of the fully fine-tuned GDINO+MedSAM baseline, and on 12 real German dictations it extracts the correct tumor class in 11 of 12 cases. The intended upshot is that dense pixel annotations and full model fine-tuning are not prerequisites for clinically useful segment

What carries the argument

The load-bearing mechanism is the detection-driven prompt chain. A negation-aware NLP stage turns transcribed dictation into a short class prompt (e.g., 'glioma'); a LoRA-augmented Grounding DINO detector (low-rank adaptation inserted in attention, cross-attention, feed-forward, and text-encoder layers) aligns that prompt to image regions and outputs bounding boxes; a frozen MedSAM turns each box into a pixel mask. LoRA is the parameter-efficiency device: fewer than 5% of model weights are updated, while the pretrained cross-modal knowledge stays fixed.

Load-bearing premise

The load-bearing premise is that healthy MRIs will not routinely produce tumor masks; the paper's qualitative note that on healthy scans the detector flags the entire brain region, with no false-positive metrics reported, leaves this assumption unquantified.

What would settle it

Count the fraction of healthy BRISC 2025 test images for which LoGSAM returns a non-empty mask; a high fraction would directly contradict the tumor-segmentation utility claimed for the pipeline.

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

If this is right

New segmentation tasks could be tackled with box-level annotations and frozen foundation models, reducing the need for dense pixel labels.
Non-English dictation can drive segmentation: the German-dictation evaluation yields correct tumor-class extraction in 11 of 12 unseen cases.
Out-of-distribution detection remains serviceable: LoRA-augmented GDINO retains 0.7955 mAP@50 on an unseen brain-MRI set, enough for MedSAM to produce usable masks.
The localization and segmentation stages are decoupled, so either module can be swapped without retraining the other.
The speech-to-prompt stage can suppress many false positives through negation-aware NLP, which matters because radiology reports routinely contain negated findings.

Where Pith is reading between the lines

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

If healthy MRIs often trigger whole-brain masks (as the qualitative note suggests), clinical use would require a no-tumor gate; measuring specificity on healthy slices is the natural next experiment.
The same chain may transfer to other text-guided imaging tasks—chest X-ray findings, ultrasound organ localization—wherever a grounding model can align report phrases to image regions.
End-to-end Dice from the German dictations, not just class-extraction accuracy, would strengthen the speech-to-mask claim; the paper reports class accuracy separately from segmentation Dice.
The LoRA rank sweet spot (r=64, with degradation at r=128) hints that low-rank adaptation has a capacity ceiling for grounding models, which could inform parameter budgets in other medical domains.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A sensible modular speech-to-segmentation pipeline with real parameter-efficiency numbers, but the headline Dice is measured with clean text prompts, not with the speech-derived prompts the system is named for.

read the letter

LoGSAM is a competent integration of existing pieces: Whisper transcription, negation-aware spaCy/negspaCy extraction, LoRA-tuned Grounding DINO, and frozen MedSAM. The individual components are known, and refs [5] and [12] already pair LoRA-tuned grounding with SAM-family segmentation, so the new bit is the speech-to-prompt bridge and the brain-MRI application. That is a legitimate extension, and the paper does several things well: the LoRA rank and injection-location ablations are clean; the OOD detection numbers give some sense of generalization; and the authors show a concrete ASR failure case instead of sweeping it under the rug. The GitHub link is a plus.

The main result holds for text-prompted localization: LoRA-GDINO plus frozen MedSAM reaches 0.8032 Dice, about 98.6% of the fully fine-tuned baseline at under 5% trainable parameters. That is a useful trade-off and worth reporting.

The paper's weak spot is the gap between that result and the speech-to-segmentation claim. The 80.32% Dice is obtained with clean text prompts. The German dictation experiment is only scored as class extraction, 11/12 cases; no masks, no Dice, no end-to-end output is reported. One dictation already failed at the ASR stage, so prompt errors do occur, and their effect on GDINO and MedSAM is unknown. The abstract's wording that LoGSAM 'attains a Dice score of 80.32%' overstates what was measured. I would not call this a fatal flaw as the modular design is plausible and the components are individually validated, but the integrated claim needs direct evidence.

Also worth fixing: the 80.32% is labeled state-of-the-art while the paper's own fully fine-tuned baseline is 0.8145; that should be reworded. The healthy-MRI behavior is described as the detector identifying the whole brain, yet no specificity or false-positive metrics are reported, which matters for real use. And there is a small inconsistency: the contributions say 554 OOD Kaggle cases while the experimental section says 509 test images.

Who should read it: people working on parameter-efficient medical segmentation or multimodal clinical pipelines. It deserves a serious referee, and I would engage with it after the authors tighten the claims and add the missing end-to-end evaluation.

Referee Report

4 major / 6 minor

Summary. The paper proposes LoGSAM, a modular pipeline that converts radiologist dictation into text prompts via Whisper ASR and spaCy/negspaCy NLP, uses a LoRA-adapted Grounding DINO to localize brain tumors from these prompts, and feeds the predicted bounding boxes to a frozen MedSAM for pixel-level segmentation. On the BRISC 2025 test set, the text-prompted pipeline attains a mean Dice of 0.8032, reaching 98.6% of a fully fine-tuned GDINO+MedSAM baseline while updating under 5% of parameters. The same pipeline is evaluated on 12 German dictations, where the class-extraction stage is correct in 11/12 cases. Detection performance is also reported on an out-of-distribution Kaggle BBOX dataset.

Significance. If the full speech-to-segmentation chain were validated end-to-end, the paper would be a useful demonstration that large pretrained vision-language and segmentation models can be assembled with minimal trainable parameters for a clinically oriented task. The public code, the modular decomposition, and the explicit LoRA ablation are strengths. However, the current evidence supports a modular proof-of-concept rather than the claimed integrated speech-to-segmentation system, and the absent healthy-case metrics materially weaken the clinical utility claim.

major comments (4)

[§3, Tables 1 and 3] The headline Dice result of 0.8032 is computed with clean text prompts, not with prompts produced by the speech/NLP module. The German-dictation experiment is reported only as 11/12 class-extraction accuracy, with no bounding boxes or segmentation masks evaluated on those 12 cases. Because the abstract calls LoGSAM a speech-to-segmentation pipeline and reports the 80.32% Dice as its score, the integrated claim is not directly tested. The authors should provide end-to-end Dice, or at least localization box quality, on the dictation-driven cases; alternatively, they should explicitly state that the segmentation score is measured under clean-text prompting.
[§3, Qualitative Assessment] The paper acknowledges that for healthy MRIs 'the detection head identifies the entire brain region,' yet no healthy-case metrics (false-positive rate, specificity, or per-class Dice) are reported. Since healthy images constitute ~20% of BRISC and healthy dictations are interpreted via default negation, a system that regularly masks the whole brain in healthy scans would not be clinically useful. The authors should quantify how often healthy slices receive tumor masks and report the corresponding segmentation-level false-positive behavior.
[Table 3 and Abstract] The 'state-of-the-art' label for the 0.8032 Dice is unsupported as written. The only comparator in Table 3 is Trust-Refined U-Net (0.7980) and an in-house fully fine-tuned GDINO+MedSAM baseline (0.8145); no external BRISC leaderboard comparison, confidence intervals, or significance tests are provided. The authors should either provide a rigorous comparison with existing BRISC methods and statistical uncertainty or temper the claim to 'competitive with a fully fine-tuned baseline.'
[Table 2, OOD evaluation] The out-of-distribution evaluation is limited to detection on Kaggle BBOX and does not include segmentation results on that dataset. The LoRA-augmented model drops from 0.8428 to 0.7955 mAP@50 relative to the fully fine-tuned baseline in the OOD setting, which is a larger degradation than in-distribution. The claim of 'robust generalization' should be scoped to detection and should explicitly acknowledge this gap, ideally with confidence intervals and an analysis of failure cases.

minor comments (6)

[§3, first paragraph of Detection-to-Segmentation] Typos: 'performstronglyonBRISC' and 'introducesamodest reduction' should be separated.
[Table 1 caption and text] The text says the speech module was tested on 12 cases, but Table 1 shows only 6; this should be made consistent and the full confusion matrix or per-case outcomes for all 12 cases should be shown.
[§3, Speech-to-prompt generation] The reported WER of 5.7% is cited from the Whisper model's general benchmark, not measured on the 12 radiologist dictations. The authors should report WER on their own test dictations.
[§3, Speech-to-prompt generation] The 11/12 success rate on n=12 has a wide confidence interval; reporting only a raw percentage overstates precision.
[Abstract] The phrase 'full pipeline using German dictations ... achieving 91.7% case-level accuracy' should be clarified as case-level class-extraction accuracy, not end-to-end segmentation accuracy.
[§2.2, Implementation Details] The authors mention 'negspaCy gating mechanism' but the reference list cites the original NegEx paper [18]; the distinction between the implemented component and the original algorithm should be stated explicitly.

Circularity Check

0 steps flagged

No significant circularity: LoGSAM's headline Dice is an empirical test-set measurement with train/test separation; the speech stage is separately evaluated and no load-bearing self-citation or definitional reduction appears.

full rationale

The paper's central derivation is an empirical pipeline evaluation, not a definitional or self-referential chain. The headline result (mean Dice 0.8032 on BRISC 2025, Table 3) is obtained by training LoRA-augmented GDINO on BRISC bounding-box annotations and then measuring segmentation quality on the BRISC test split with frozen MedSAM. This is a standard train/test protocol: the LoRA parameters are fitted to training data, while the reported Dice is computed on held-out test images. The confidence threshold tau=0.3 and LoRA rank r=64 are ordinary hyperparameters selected by ablation; neither defines the target metric, nor is the Dice score a renamed version of any fitted quantity. The speech-to-prompt module is evaluated separately on 12 German dictations (Table 1, 11/12 case-level class-extraction accuracy), and this evaluation does not presuppose the segmentation scores; conversely, the Dice measurement uses clean textual prompts. The fact that the two stages are not evaluated end-to-end in a single numeric metric is an evaluation-coverage gap, not circularity. There is no load-bearing self-citation: references to Whisper, spaCy, negspaCy, GDINO, and MedSAM point to external pretrained systems, and no 'uniqueness theorem' or prior author result is invoked to force the design. The qualitative statement that for healthy MRIs 'the detection head identifies the entire brain region' identifies a potential false-positive risk, but this is a correctness/validation concern, not a circular-reasoning concern. Because the main claims are supported by an independent empirical benchmark with a train/test split and an OOD evaluation set, the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central numbers depend on the reliability of pretrained speech, NLP, detection, and segmentation models, on hand-built vocabulary and negation rules, and on the correctness of public dataset annotations; none of these are derived in this paper.

free parameters (3)

LoRA rank = r=64
Selected by ablation (Table 4) as the best accuracy/parameter trade-off; the headline 4.96% trainable-parameter figure depends on it.
Detection confidence threshold τ = 0.3
Chosen by hand to filter Grounding DINO boxes before MedSAM; directly sets which boxes become masks and thus the Dice score.
LoRA training schedule = 125 epochs, AdamW lr=2e-4, OneCycle 10% warmup, gradient clip 5.0
Set for the BRISC adaptation; no sensitivity analysis is reported, so the result is conditioned on this schedule.

axioms (4)

domain assumption Pretrained component reliability
Section 2.2 relies on Whisper large-v3's reported 5.7% WER, spaCy's 92.6% entity accuracy, GDINO's text-conditioned localization, and MedSAM's prompt-based segmentation without verifying each on this domain beyond the reported results.
domain assumption Curated clinical vocabulary coverage
The synonym dictionary and negation triggers are hand-curated, and the paper's own results show one of 12 dictations fails on an out-of-vocabulary transcription ('glioblastoma' -> 'glioplaston'), so coverage is load-bearing but only tested on 12 cases.
domain assumption Dataset label reliability
BRISC 2025 and Kaggle BBOX annotations are assumed correct for training and evaluation; the paper notes BRISC lacks acquisition metadata and patient overlap is only manually minimized.
domain assumption Healthy-case behavior is benign
The paper assumes that healthy MRIs not producing tumor masks is acceptable, but reports the detector identifies the entire brain region on healthy cases; this assumption is unverified quantitatively.

pith-pipeline@v1.3.0-alltime-deepseek · 6608 in / 14500 out tokens · 139429 ms · 2026-08-02T17:55:56.277126+00:00 · methodology

0 comments

read the original abstract

Precise localization and delineation of brain tumors using magnetic resonance imaging (MRI) are essential for planning therapy and guiding surgical decisions. To address this, we propose LoGSAM, a parameter-efficient, detection-driven framework that transforms radiologist dictation into text prompts for foundation-model-based localization and segmentation. Radiologist speech is first transcribed and translated using a pretrained Whisper ASR model, followed by negation-aware clinical NLP to extract tumor-specific textual prompts. These prompts guide text-conditioned tumor localization via a LoRA-adapted vision-language detection model, Grounding DINO (GDINO). The predicted bounding boxes are used as prompts for MedSAM to generate pixel-level tumor masks without any additional fine-tuning. On BRISC 2025, LoGSAM attains a Dice score of 80.32%, reaching 98.6% of a fully fine-tuned GDINO + MedSAM baseline while training fewer than 5\% of its parameters, indicating a favorable accuracy/parameter trade-off. In addition, we evaluate the full pipeline using German dictations from a board-certified radiologist on unseen MRI scans, achieving 91.7\% case-level class-extraction accuracy. These results highlight the feasibility of constructing a modular speech-to-segmentation pipeline from pretrained foundation models with minimal parameter updates.

Figures

Figures reproduced from arXiv: 2603.17576 by Andreas Maier, Melika Qahqaie, Mohammad Robaitul Islam Bhuiyan, Paula Andrea P\'erez-Toro, Sheethal Bhat, Tom\'as Arias-Vergara, Tri-Thien Nguyen.

**Figure 2.** Figure 2: Qualitative results at different stages of the proposed detec [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 1 linked inside Pith

[1]

In: European Conference on Computer Vision (ECCV), pp

Liu, S., Zeng, Z., Ren, T., et al.: Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. In: European Conference on Computer Vision (ECCV), pp. 38–55. Springer, Heidelberg (2024)

2024
[2]

In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 234–241 (2015)

2015
[3]

W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust Speech Recognition via Large-Scale Weak Supervision

Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust Speech Recognition via Large-Scale Weak Supervision. In: Int. Conference on Machine Learning (ICML), pp. 28492–28518. PMLR (2023)

2023
[4]

Nature Communications,15(1), 654 (2024)

Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications,15(1), 654 (2024)

2024
[5]

IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control (2025)

Rasaee, H., Koleilat, T., Rivaz, H.: Grounding DINO-US-SAM: Text-Prompted Multi-Organ Segmentation in Ultrasound with LoRA-Tuned Vision-Language Mod- els. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control (2025)

2025
[6]

J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (ICLR),1(2), 3 (2022)

2022
[7]

Acta Neuropathologica 131(6), 803–820 (2016)

Louis,D.N., Perry, A.,Reifenberger,G., etal.:The 2016World HealthOrganization Classification of Tumors of the Central Nervous System. Acta Neuropathologica 131(6), 803–820 (2016)

2016
[8]

Physics in Medicine & Biology58(13), R97 (2013)

Bauer, S., Wiest, R., Nolte, L.-P., Reyes, M.: A Survey of MRI-Based Medical Image Analysis for Brain Tumor Studies. Physics in Medicine & Biology58(13), R97 (2013)

2013
[9]

Frontiers in Oncology15, 1643208 (2025)

Han, W., Dong, X., Wang, G., Ding, Y., Yang, A.: Application and improvement of YOLO11 for brain tumor detection in medical images. Frontiers in Oncology15, 1643208 (2025)

2025
[10]

arXiv preprint arXiv:1906.07155 (2019)

Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)

Pith/arXiv arXiv 1906
[11]

Insights into Imaging14(1), 47 (2023)

Jörg, T., Kämpgen, B., Feiler, D., Müller, L., Düber, C., Mildenberger, P., Jung- mann, F.: Efficient structured reporting in radiology using an intelligent dialogue system based on speech recognition and natural language processing. Insights into Imaging14(1), 47 (2023)

2023
[12]

In: International Conference on Medical Imaging and Computer-Aided Diagnosis, pp

Hu, J., Xu, X., Zou, Z.: LoRA-MedSAM: Efficient medical image segmentation. In: International Conference on Medical Imaging and Computer-Aided Diagnosis, pp. 154–164. Springer, Singapore (2024)

2024
[13]

L., et al.: Radiotherapy Planning and Evaluation

Chang, E. L., et al.: Radiotherapy Planning and Evaluation. Neurosurgery Clinics of North America18(1), 43–53 (2007)

2007
[14]

D., Langlotz, C

Zhang, Y., Jiang, H., Miura, Y., Manning, C. D., Langlotz, C. P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare Conference, pp. 2–25. PMLR (2022)

2022
[15]

In: Proceedings of the Conference on Empirical Methods in NLP (EMNLP), vol

Wang, Z., Wu, Z., Agarwal, D., Sun, J.: MedCLIP: Contrastive learning from unpaired medical images and text. In: Proceedings of the Conference on Empirical Methods in NLP (EMNLP), vol. 2022, p. 3876 (2022)

2022
[16]

K., Mahala, S., Kumar, Y., Shekhawat, N

Bhamboo, A. K., Mahala, S., Kumar, Y., Shekhawat, N. S., Sharma, K.: Trust- Refined U-Net Ensemble for Uncertainty-Aware Brain Tumor Segmentation. Au- thorea Preprints (2025) 10 M.R.I. Bhuiyan et. al

2025
[17]

C., Lo, W.-Y., Dollár, P.: Segment Anything

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., Dollár, P.: Segment Anything. In: Proceed- ings of the IEEE/CVF Int. Conference on Computer Vision (ICCV), pp. 4015–4026 (2023)

2023
[18]

W., Bridewell, W., Hanbury, P., Cooper, G

Chapman, W. W., Bridewell, W., Hanbury, P., Cooper, G. F., Buchanan, B. G.: A simple algorithm for identifying negated findings and diseases in discharge sum- maries. Journal of Biomedical Informatics34(5), 301–310 (2001)

2001
[19]

Zenodo (2020)

Honnibal, M., Montani, I., Van Landeghem, S., Boyd, A.: spaCy: Industrial- Strength Natural Language Processing in Python. Zenodo (2020)

2020
[20]

Scientific Data (2026)

Fateh,A.,Rezvani,Y.,Moayedi,S.,Rezvani,S.,Fateh,F.,Fateh,M.,Abolghasemi, V.: BRISC: Annotated dataset for brain tumor segmentation and classification. Scientific Data (2026)

2026
[21]

MRI for Brain Tumor with Bounding Boxes (Kag- gle dataset),https://www.kaggle.com/datasets/ahmedsorour1/ mri-for-brain-tumor-with-bounding-boxes, last accessed 2026/02/07

2026

[1] [1]

In: European Conference on Computer Vision (ECCV), pp

Liu, S., Zeng, Z., Ren, T., et al.: Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. In: European Conference on Computer Vision (ECCV), pp. 38–55. Springer, Heidelberg (2024)

2024

[2] [2]

In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomed- ical Image Segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 234–241 (2015)

2015

[3] [3]

W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust Speech Recognition via Large-Scale Weak Supervision

Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., Sutskever, I.: Ro- bust Speech Recognition via Large-Scale Weak Supervision. In: Int. Conference on Machine Learning (ICML), pp. 28492–28518. PMLR (2023)

2023

[4] [4]

Nature Communications,15(1), 654 (2024)

Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications,15(1), 654 (2024)

2024

[5] [5]

IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control (2025)

Rasaee, H., Koleilat, T., Rivaz, H.: Grounding DINO-US-SAM: Text-Prompted Multi-Organ Segmentation in Ultrasound with LoRA-Tuned Vision-Language Mod- els. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control (2025)

2025

[6] [6]

J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models

Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Conference on Learning Representations (ICLR),1(2), 3 (2022)

2022

[7] [7]

Acta Neuropathologica 131(6), 803–820 (2016)

Louis,D.N., Perry, A.,Reifenberger,G., etal.:The 2016World HealthOrganization Classification of Tumors of the Central Nervous System. Acta Neuropathologica 131(6), 803–820 (2016)

2016

[8] [8]

Physics in Medicine & Biology58(13), R97 (2013)

Bauer, S., Wiest, R., Nolte, L.-P., Reyes, M.: A Survey of MRI-Based Medical Image Analysis for Brain Tumor Studies. Physics in Medicine & Biology58(13), R97 (2013)

2013

[9] [9]

Frontiers in Oncology15, 1643208 (2025)

Han, W., Dong, X., Wang, G., Ding, Y., Yang, A.: Application and improvement of YOLO11 for brain tumor detection in medical images. Frontiers in Oncology15, 1643208 (2025)

2025

[10] [10]

arXiv preprint arXiv:1906.07155 (2019)

Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019)

Pith/arXiv arXiv 1906

[11] [11]

Insights into Imaging14(1), 47 (2023)

Jörg, T., Kämpgen, B., Feiler, D., Müller, L., Düber, C., Mildenberger, P., Jung- mann, F.: Efficient structured reporting in radiology using an intelligent dialogue system based on speech recognition and natural language processing. Insights into Imaging14(1), 47 (2023)

2023

[12] [12]

In: International Conference on Medical Imaging and Computer-Aided Diagnosis, pp

Hu, J., Xu, X., Zou, Z.: LoRA-MedSAM: Efficient medical image segmentation. In: International Conference on Medical Imaging and Computer-Aided Diagnosis, pp. 154–164. Springer, Singapore (2024)

2024

[13] [13]

L., et al.: Radiotherapy Planning and Evaluation

Chang, E. L., et al.: Radiotherapy Planning and Evaluation. Neurosurgery Clinics of North America18(1), 43–53 (2007)

2007

[14] [14]

D., Langlotz, C

Zhang, Y., Jiang, H., Miura, Y., Manning, C. D., Langlotz, C. P.: Contrastive learning of medical visual representations from paired images and text. In: Machine Learning for Healthcare Conference, pp. 2–25. PMLR (2022)

2022

[15] [15]

In: Proceedings of the Conference on Empirical Methods in NLP (EMNLP), vol

Wang, Z., Wu, Z., Agarwal, D., Sun, J.: MedCLIP: Contrastive learning from unpaired medical images and text. In: Proceedings of the Conference on Empirical Methods in NLP (EMNLP), vol. 2022, p. 3876 (2022)

2022

[16] [16]

K., Mahala, S., Kumar, Y., Shekhawat, N

Bhamboo, A. K., Mahala, S., Kumar, Y., Shekhawat, N. S., Sharma, K.: Trust- Refined U-Net Ensemble for Uncertainty-Aware Brain Tumor Segmentation. Au- thorea Preprints (2025) 10 M.R.I. Bhuiyan et. al

2025

[17] [17]

C., Lo, W.-Y., Dollár, P.: Segment Anything

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W.-Y., Dollár, P.: Segment Anything. In: Proceed- ings of the IEEE/CVF Int. Conference on Computer Vision (ICCV), pp. 4015–4026 (2023)

2023

[18] [18]

W., Bridewell, W., Hanbury, P., Cooper, G

Chapman, W. W., Bridewell, W., Hanbury, P., Cooper, G. F., Buchanan, B. G.: A simple algorithm for identifying negated findings and diseases in discharge sum- maries. Journal of Biomedical Informatics34(5), 301–310 (2001)

2001

[19] [19]

Zenodo (2020)

Honnibal, M., Montani, I., Van Landeghem, S., Boyd, A.: spaCy: Industrial- Strength Natural Language Processing in Python. Zenodo (2020)

2020

[20] [20]

Scientific Data (2026)

Fateh,A.,Rezvani,Y.,Moayedi,S.,Rezvani,S.,Fateh,F.,Fateh,M.,Abolghasemi, V.: BRISC: Annotated dataset for brain tumor segmentation and classification. Scientific Data (2026)

2026

[21] [21]

MRI for Brain Tumor with Bounding Boxes (Kag- gle dataset),https://www.kaggle.com/datasets/ahmedsorour1/ mri-for-brain-tumor-with-bounding-boxes, last accessed 2026/02/07

2026