Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging
Open accessAbstract
Segmenting brain tumors in multi-parametric magnetic resonance imaging enables performing quantitative analysis in support of clinical trials and personalized patient care. This analysis provides the potential to impact clinical decision-making processes, including diagnosis and prognosis. In 2023, the well-established Brain Tumor Segmentation (BraTS) challenge presented a substantial expansion with eight tasks and 4,500 brain tumor cases. In this paper, we present a deep learning-based ensemble strategy that is evaluated for newly included tumor cases in three tasks: pediatric brain tumors (PED), intracranial meningioma (MEN), and brain metastases (MET). In particular, we ensemble outputs from state-of-the-art nnU-Net and Swin UNETR models on a region-wise basis. Furthermore, we implemented a targeted post-processing strategy based on a cross-validated threshold search to improve the segmentation results for tumor sub-regions. The evaluation of our proposed method on unseen test cases for the three tasks resulted in lesion-wise Dice scores for PED: 0.653, 0.809, 0.826; MEN: 0.876, 0.867, 0.849; and MET: 0.555, 0.6, 0.58; for the enhancing tumor, tumor core, and whole tumor, respectively. Our method was ranked first for PED, third for MEN, and fourth for MET, respectively.
AI snapshot
Written from the paper’s full text.
TL;DR
This paper presents a deep learning ensemble of nnU-Net and Swin UNETR with task-specific post-processing for brain tumor segmentation in the BraTS 2023 pediatric, meningioma, and metastasis tasks. It ranked first for PED, third for MEN, and fourth for MET on unseen test cases.
Key findings
- On unseen test cases, lesion-wise Dice scores were PED: 0.653, 0.809, 0.826; MEN: 0.876, 0.867, 0.849; and MET: 0.555, 0.6, 0.58 for enhancing tumor, tumor core, and whole tumor, respectively.
- The method ranked first for PED, third for MEN, and fourth for MET.
- The combination of ensemble models and post-processing yielded clear improvements across PED, MEN, and MET, with the improvement more noticeable in pediatric tumor segmentation.
- ET identification was the most challenging in PED, as shown by the lowest Dice coefficient for that region.
Methods
- The authors ensembled nnU-Net and Swin UNETR outputs on a label-wise/region-wise basis, using five-fold cross-validation; for MET, only nnU-Net was used after experimental evaluation because Swin UNETR showed inferior performance.
- Data were multi-parametric MRI including T1, T1CE, T2, and T2-FLAIR. Splits were PED: training n=99, validation n=45; MEN: training n=1,000, validation n=141; MET: training n=165, validation n=31, with additional testing datasets.
- Post-processing removed disconnected regions smaller than 130 voxels for PED, 110 for MEN, and 15 for MET, and applied PED-specific label redefinition using cross-validated thresholds: ET/WT 0.04 and ED/WT 1.00.
Limitations
- The paper states that PED and MET were conducted with limited available data, and that each task’s training was restricted to task-specific data per BraTS 2023 guidelines.
- The authors state that additional pre-training outside the challenge could improve generalization, and they suggest self-supervised learning, foundation models, or generative AI to augment limited training sets.
- Evident from the text: the ensemble did not use Swin UNETR for MET because it underperformed, so the final MET solution relied on nnU-Net alone.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
6 of 7 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured
- Findingown data
For the enhancing tumor (ET) label, the optimal post-processing threshold was 0.04.
“For ET, we obtained an optimal threshold of 0.04.”
Quote in the paperNot a substantial claim - Findingown data
For the MEN and MET tasks, a better balance between label segmentation was achieved.
“For the MEN and MET tasks, a better balance between label segmentation was achieved.”
Quote in the paperSubstantial claim - Findingown data
For the edematous area (ED) label, the optimal post-processing threshold was 1.00.
“On the other hand, Fig. 5b displays the lesion-wise Dice score vs ED/WT threshold curve obtained from the same process but applied to the ED label, which yielded an optimal threshold of 1.00.”
Quote in the paperNot a substantial claim - Interpretationnone given
The proposed strategy minimized undesired smaller contributions that would negatively impact lesion-wise performance.
“Our strategy minimized undesired smaller contributions, which would negatively impact lesion-wise performance.”
Quote in the paperSubstantial claim- Rests on an assumption: The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
- Interpretationown data
The ensemble prediction from nnU-Net and Swin UNETR provided better segmentation accuracy for all tumor types.
“The ensemble prediction from nnU-net and Swin UNETR provided better segmentation accuracy for all tumor types.”
Quote in the paperSubstantial claim
- Interpretationown data
In the PED task, enhancing tumor (ET) identification was the most challenging sub-region, as evidenced by the lowest Dice coefficient for this region.
“Among these three sub-regions, ET identification was the most challenging in the PED task, as evidenced by the lowest Dice coefficient for this region.”
Quote in the paperSubstantial claim - Interpretationown data
The combination of ensemble models and post-processing procedures yielded clear improvements in performance for the PED, MEN, and MET tasks.
“Overall, the combination of ensemble models and post-processing procedures yielded clear improvements in the performance for the PED, MEN, and MET tasks.”
Quote in the paperSubstantial claim- Rests on an assumption: The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
- Premisenone given
Lesion-centered metrics result in severe penalties for isolated volumes that do not correspond to actual lesions.
“Moreover, the novel metrics centered on lesions result in severe penalties for isolated volumes that do not correspond to actual lesions.”
Quote in the paperSubstantial claim- Rests on an assumption: The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
- May combine two claims. 9
- Classified as probably not a substantial claim. 5
- The quote only partly states this claim. 1
May combine two claims.
PremisecitationSwin UNETR's local window self-attention outperforms traditional ViT and is well suited to multiscale tasks.
“Swin UNETR’s innovative local window self-attention outperforms traditional ViT, which is well suited to multiscale tasks.”
Quote in the paperNot a substantial claim3 of 3 sources checked- cites [11] Hatamizadeh, A., Nath, V., Tang, Y., et al.: Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. In: Crimi, A., Bakas, S. (eds.) Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries, pp. 272–284. Springer, Cham (2022). https://doi.org/10.1007/978-3-031-08999-2_22 source
- cites [12] Hatamizadeh, A., Tang, Y., Nath, V., et al.: UNETR: transformers for 3D medical image segmentation. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 574–584 (2022) source
The Implementation Details (p. 5) describe UNETR’s encoder as following ViT-B16, while the Discussion (p. 7) reports UNETR’s performance. Our copy does not report Swin UNETR’s local-window self-attention, a comparison with traditional ViT, or its suitability for multiscale tasks (searched: Swin UNETR, local window self-attention, multiscale, ViT). (LLM, conf=0.76)
Experimental — accuracy not yet measured - cites [27] Tang, Y., Yang, D., Li, W., Roth, H.R., et al.: Self-supervised pre-training of swin transformers for 3D medical image analysis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20730–20740 (2022) source
May combine two claims.
PremisecitationAlthough a few previous works addressed pediatric brain tumors, their automatic segmentation remains challenging due to a lack of training data and variation in heterogeneous histologic sub-regions.
“Although, a few previous works addressed pediatric brain tumors [10,14,20], their automatic segmentation remains challenging. This is due to a lack of training data and variation in heterogeneous histologic sub-regions including peritumoral edematous/invaded tissue, necrotic core, and enhancing tumor.”
Quote in the paperSubstantial claim2 of 3 sources checked- cites [10] Fathi Kazerooni, A., Arif, S., Madhogarhia, R., et al.: Automated tumor segmentation and brain tissue extraction from multiparametric MRI of pediatric brain tumors: a multi-institutional study. Neuro-Oncol. Adv. 5(1), vdad027 (2023) source
The introductory passage (labeled Abstract > Conclusions; page unavailable) describes a few earlier pediatric segmentation studies and limitations in their performance and ability to distinguish subregions. The Methods and Results (pages unavailable) also describe difficulty distinguishing enhancing tumor, non-enhancing tumor, and edema; the Discussion links limited edema examples to class imbalance. The paper does not establish a general lack of training data as the cause of the challenge, and its listed subregions do not include the claim’s necrotic core or invaded tissue. (LLM, conf=0.84)
Experimental — accuracy not yet measured - cites [14] Jiang, Z., Parida, A., Anwar, S.M., et al.: Automatic visual acuity loss prediction in children with optic pathway gliomas using magnetic resonance imaging. In: 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 1–5 (2023) source
The Introduction says automatic segmentation of the anterior visual pathway in children with optic pathway gliomas is difficult and that obtaining sufficient training data for rare pediatric brain tumors is challenging. The Methods list necrotic, peritumoral edematous, and enhancing regions as labels in an adult glioblastoma dataset; the paper does not say that variation among those regions causes the pediatric segmentation difficulty. (LLM, conf=0.87)
Experimental — accuracy not yet measured - cites [20] Liu, X., Bonner, E., Jiang, Z., et al.: Automatic segmentation of rare pediatric brain tumors using knowledge transfer from adult data. In: 20th IEEE International Symposium on Biomedical Imaging (2023) source
May combine two claims.
PremisecitationThe nnU-Net framework autonomously adjusts its internal configurations according to the specific imaging modality and unique attributes of each dataset, resulting in improved segmentation performance and generalization compared to other state-of-the-art methods for biomedical image segmentation.
“According to the specific imaging modality and unique attributes of each dataset, the framework autonomously adjusts its internal configurations [13]. This results in an improved segmentation performance and generalization when compared to other state-of-the-art methods for biomedical image segmentation.”
Quote in the paperSubstantial claim1 of 1 sources checked- cites [13] Isensee, F., Jaeger, P.F., Kohl, S.A., et al.: nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods 18(2), 203–211 (2021) source
May combine two claims.
Findingown dataFor the MET task, Swin UNETR showed inferior performance compared to nnU-Net, so only nnU-Net was used.
“Finally, for the MET task, only nnU-Net, trained on all labels, was used after experimental evaluation, as Swin UNETR showed inferior performance in this scenario.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataOn unseen test cases for the BraTS 2023 PED, MEN, and MET tasks, the proposed method achieved lesion-wise Dice scores for PED of 0.653, 0.809, and 0.826; for MEN of 0.876, 0.867, and 0.849; and for MET of 0.555, 0.6, and 0.58, for enhancing tumor, tumor core, and whole tumor, respectively.
“The evaluation of our proposed method on unseen test cases for the three tasks resulted in lesion-wise Dice scores for PED: 0.653, 0.809, 0.826; MEN: 0.876, 0.867, 0.849; and MET: 0.555, 0.6, 0.58; for the enhancing tumor, tumor core, and whole tumor, respectively.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataThe proposed method was ranked first for the PED task, third for the MEN task, and fourth for the MET task.
“Our method was ranked first for PED, third for MEN, and fourth for MET, respectively.”
- Also stated in: § Conclusion · boosted
“The success of our method is demonstrated by the challenge ranking positions on unseen test cases, being the winner in PED task, ranked third for MEN, and fourth for MET task.”
Quote in the paperSubstantial claim- Also stated in: § Conclusion · boosted
Classified as probably not a substantial claim. May combine two claims. The quote only partly states this claim.
Study design and conductown dataFor each of the three tasks, both the nnU-Net and Swin UNETR models were trained using a five-fold cross-validation approach.
“For each of the three tasks, we trained a full-resolution 3D nnU-Net (v2) model using a five-fold cross-validation approach.”
- Also stated in: § Deep Learning Models and Experimental Setup · plain
“For each of the three tasks, we trained a full-resolution 3D Swin UNETR model using a five-fold cross-validation approach.”
Quote in the paperSubstantial claim- Also stated in: § Deep Learning Models and Experimental Setup · plain
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductnone givenThe performance evaluation on the validation datasets was performed automatically by the challenge's digital platform, with no access to the validation ground truth data.
“This performance evaluation was performed automatically by the challenge’s digital platform, with no access to the validation ground truth data.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Premisenone givenBinary metrics disproportionately penalize false positive predictions.
“Consequently, binary metrics disproportionately penalize false positive predictions.”
Quote in the paperNot a substantial claim- Rests on an assumption: The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenEach task's training process was restricted to the usage of task-specific data, adhering to the predefined guidelines of the BraTS 2023 challenge.
“Each task’s training process was restricted to the usage of task-specific data, adhering to the predefined guidelines of the BraTS 2023 challenge.”
Quote in the paperNot a substantial claimMay combine two claims.
InterpretationreasoningThe targeted post-processing strategy based on cross-validated thresholding search improved performance in all tasks, with the improvement being more noticeable in pediatric tumor segmentation.
“Furthermore, our targeted post-processing strategy based on cross-validated thresholding search improved the performance in all tasks, being more noticeable in pediatric tumor segmentation due to the label-wise targeted post-processing steps.”
Quote in the paperSubstantial claim- Rests on an assumption: The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
Classified as probably not a substantial claim.
Study design and conductown dataThe optimal post-processing threshold values were determined by experimenting with the cross-validation data.
“These optimal threshold values were determined by experimenting with the cross-validation data.”
- Also stated in: § Post-processing · plain
“These optimal threshold values were also determined by experimenting with the cross-validation data.”
Quote in the paperSubstantial claim- Also stated in: § Post-processing · plain
The paper relies on this without support.
The BraTS 2023 challenge introduced new performance metrics that assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).
“New performance metrics were introduced in this year’s BraTS challenge to assess segmentation models at a lesion-wise level rather than over the entire/multiple tumor region(s).”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Conclusion credits an ensemble for MET, but MET used only nnU-Net
MajorThe conclusion attributes better segmentation accuracy for all tumor types to an ensemble of nnU-Net and Swin UNETR. The methods state that the MET task used only nnU-Net, so its results cannot demonstrate a benefit from that ensemble.
“The ensemble prediction from nnU-net and Swin UNETR provided better segmentation accuracy for all tumor types.”
“Finally, for the MET task, only nnU-Net, trained on all labels, was used after experimental evaluation, as Swin UNETR showed inferior performance in this scenario.”
Cites
U-Net: Convolutional Networks for Biomedical Image Segmentation
Ronneberger, O., Fischer, P., & Brox, T. (2015) · Lecture notes in computer science · doi:10.1007/978-3-319-24574-4_28 · Cited by 99,404 · cited 1× · see citing works
“nnU-Net The nnU-Net, which is based on the U-Net architecture [26], is a self-configuring deep learning framework for semantic segmentation.”
MethodAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 21,793 · cited 1× · see citing works
“Swin UNETR The Swin UNETR framework employs a vision transformer (ViT)- based [9] hierarchical structure for localized self-attention using non-overlapping windows [11,12,27].”
MethodnnU-Net: a self-configuring method for deep learning-based biomedical image segmentation
Isensee, F., Jaeger, P. F., Kohl, S. A. A., Petersen, J., & Maier-Hein, K. H. (2021) · Nature Methods · doi:10.1038/s41592-020-01008-z · Cited by 10,477 · cited 3× · see citing works
“Region-based training was employed and the patch size was determined by the GPU memory allocation, favoring larger patches while remaining within the GPU’s capacity [13].”
BackgroundMethodLarian Studios Dev Team (2024) · doi:10.1145/3292500.3330701 · Cited by 8,618 · cited 1× · see citing works
“Hyper-parameter optimization was carried out using Optuna [2]: 6 .”
MethodThe multimodal brain tumor image segmentation benchmark (BRATS)
Menze, B. H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al. (2015) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2014.2377694 · Cited by 6,781 · cited 1× · see citing works
“The brain tumor segmentation (BraTS) challenge held in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) conference, established in 2012, has generated a benchmark dataset for the segmentation of adult brain gliomas [4,5,6,7,21].”
MethodUNETR: Transformers for 3D Medical Image Segmentation
Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H. R., & Xu, D. (2022) · 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) · doi:10.1109/wacv51458.2022.00181 · Cited by 3,586 · cited 1× · see citing works
“Swin UNETR The Swin UNETR framework employs a vision transformer (ViT)- based [9] hierarchical structure for localized self-attention using non-overlapping windows [11,12,27].”
MethodBakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J. S., Freymann, J. B., Farahani, K., & Davatzikos, C. (2017) · Scientific Data · doi:10.1038/sdata.2017.117 · Cited by 3,049 · cited 1× · see citing works
“The brain tumor segmentation (BraTS) challenge held in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) conference, established in 2012, has generated a benchmark dataset for the segmentation of adult brain gliomas [4,5,6,7,21].”
MethodSwin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images
Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H. R., & Xu, D. (2022) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-08999-2_22 · Cited by 2,273 · cited 1× · see citing works
“Swin UNETR The Swin UNETR framework employs a vision transformer (ViT)- based [9] hierarchical structure for localized self-attention using non-overlapping windows [11,12,27].”
MethodOptuna: A Next-generation Hyperparameter Optimization Framework
Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1907.10902 · Cited by 1,675
Self-Supervised Pre-Training of Swin Transformers for 3D Medical Image Analysis
Tang, Y., Yang, D., Li, W., Roth, H. R., Landman, B., Xu, D., Nath, V., & Hatamizadeh, A. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.02007 · Cited by 849 · cited 2× · see citing works
“Swin UNETR The Swin UNETR framework employs a vision transformer (ViT)- based [9] hierarchical structure for localized self-attention using non-overlapping windows [11,12,27].”
MethodThe SRI24 multichannel atlas of normal adult human brain structure
Torsten Rohlfing, B., Zahr, N. M., Sullivan, E. V., & Pfefferbaum, A. (2009) · Human Brain Mapping · doi:10.1002/hbm.20906 · Cited by 477 · cited 2× · see citing works
“The BraTS 2023 challenge has expanded to a cluster of challenges, encompassing a variety of tumor types alongside augmentation tasks [1,10,15,17,19,25].”
BackgroundMethodBaid, U., Ghodasara, S., Mohan, S., Bilello, M., Calabrese, E., Colak, E., Farahani, K., Kalpathy–Cramer, J., Kitamura, F., Pati, S., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.02314 · Cited by 431 · cited 1× · see citing works
“The brain tumor segmentation (BraTS) challenge held in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) conference, established in 2012, has generated a benchmark dataset for the segmentation of adult brain gliomas [4,5,6,7,21].”
MethodSegmentation Labels for the Pre-operative Scans of the TCGA-GBM collection
Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., & Davatzikos, C. (2017) · The Cancer Imaging Archive · doi:10.7937/k9/tcia.2017.klxwjj1q · Cited by 316 · see citing works
Segmentation Labels for the Pre-operative Scans of the TCGA-LGG collection
Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., & Davatzikos, C. (2017) · The Cancer Imaging Archive · doi:10.7937/k9/tcia.2017.gjq7r0ef · Cited by 290 · see citing works
Federated benchmarking of medical artificial intelligence with MedPerf
Karargyris, A., Umeton, R., Sheller, M. J., Aristizabal, A., George, J., Wuest, A., Pati, S., Kassem, H., Zenk, M., Baid, U., et al. (2023) · Nature Machine Intelligence · doi:10.1038/s42256-023-00652-2 · Cited by 188 · cited 1× · see citing works
“The BraTS 2023 challenge has expanded to a cluster of challenges, encompassing a variety of tumor types alongside augmentation tasks [1,10,15,17,19,25].”
BackgroundDeclines in Cancer Death Rates Among Children and Adolescents in the United States, 1999-2014.
Curtin, S. C., Miniño, A., & Anderson, R. N. (2016) · PubMed · Cited by 146 · see citing works
Pediatric diffuse intrinsic pontine glioma: where do we stand?
Rashed, W. M., Maher, E., Adel, M., Saber, O., & Zaghloul, M. S. (2019) · Cancer and Metastasis Reviews · doi:10.1007/s10555-019-09824-2 · Cited by 71 · cited 1× · see citing works
“For example, the median overall survival was reported to be less than one year for pediatric diffuse mid-line gliomas (DMGs, which replaced the formerly known diffuse intrinsic pontine gliomas (DIPGs) to emphasize that the disease may affect areas other than the pons) [24].”
BackgroundKazerooni, A. F., Khalili, N., Liu, X., et al. (2023) · PubMed · doi:10.48550/arxiv.2305.17033 · Cited by 61 · cited 1× · see citing works
“Particularly, these tasks include pediatric brain tumors (PED) [16], intracranial meningioma (MEN) [18], and brain metastasis (MET) [22].”
BackgroundSelf-citationFathi Kazerooni, A., Arif, S., Madhogarhia, R., Khalili, N., Haldar, D., Bagheri, S., Familiar, A. M., Anderson, H., Haldar, S., Tu, W., et al. (2023) · Neuro-Oncology Advances · doi:10.1093/noajnl/vdad027 · Cited by 55 · cited 2× · see citing works
“The BraTS 2023 challenge has expanded to a cluster of challenges, encompassing a variety of tumor types alongside augmentation tasks [1,10,15,17,19,25].”
BackgroundMoawad, A. W., Janas, A., Baid, U., et al. (2025) · The Journal of Machine Learning for Biomedical Imaging · doi:10.59275/j.melba.2025-9bd3 · Cited by 40 · see citing works
Self-citationBrain tumor segmentation in Sub-Saharan Africa patient population: The BraTS-Africa challenge
Adewole, M., Rudie, J. D., Gbadamosi, A., Zhang, D., Raymond, C., Toyobo, O. O., Omidiji, O. A. T., Akinola, R., Suwaid, M. A., Daji, F. Y., et al. (2026) · Neuro-Oncology Advances · doi:10.1093/noajnl/vdag082 · Cited by 39 · see citing works
Self-citationThe ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge 2023: Intracranial MeningiomaPreprint
LaBella, D., Adewole, M., Alonso‐Basanta, M., et al. (2023) · PubMed · doi:10.48550/arxiv.2305.07642 · Cited by 27 · cited 2× · see citing works
“Particularly, these tasks include pediatric brain tumors (PED) [16], intracranial meningioma (MEN) [18], and brain metastasis (MET) [22].”
BackgroundSelf-citationAutomatic Segmentation of Rare Pediatric Brain Tumors Using Knowledge Transfer From Adult Data
Liu, X., Bonner, E. R., Jiang, Z., Roth, H. R., Muhammad Anwar, S., Packer, R. J., Bornhorst, M., & Linguraru, M. G. (2023) · 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi53787.2023.10230757 · Cited by 24 · cited 1× · see citing works
“Although, a few previous works addressed pediatric brain tumors [10,14,20], their automatic segmentation remains challenging.”
BackgroundSelf-citationLi, H., Conte, G. M., Hu, Q., et al. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2305.09011 · Cited by 12 · cited 1× · see citing works
“The BraTS 2023 challenge has expanded to a cluster of challenges, encompassing a variety of tumor types alongside augmentation tasks [1,10,15,17,19,25].”
BackgroundSelf-citationJiang, Z., Parida, A., Anwar, S. M., Tang, Y., Roth, H. R., Fisher, M. J., Packer, R. J., Avery, R. A., & Linguraru, M. G. (2023) · 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) · doi:10.1109/embc40787.2023.10339961 · Cited by 8 · cited 1× · see citing works
“Although, a few previous works addressed pediatric brain tumors [10,14,20], their automatic segmentation remains challenging.”
BackgroundSelf-citationSPCXR: Self-supervised Pretraining using Chest X-rays Towards a Domain Specific Foundation ModelPreprint
Anwar, S. M., Parida, A., Atito, S., Awais, M., Niño, G., Kitler, J., & Linguraru, M. G. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2211.12944 · Cited by 6 · cited 1× · see citing works
“Towards this, we argue that self-supervised learning strategies and the concept of foundation model [3,23] holds the potential to augment model’s generalizability, thus making it suitable for a more extensive range of clinical applications.”
BackgroundSelf-citationKofler, F., Meissen, F., Steinbauer, F., et al. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2305.08992 · Cited by 5 · see citing works
Self-citationMoawad, A. W., Janas, A., Baid, U., Ramakrishnan, D. S., Saluja, R., Ashraf, N., Maleki, N., Jekel, L., Yordanov, N., Fehringer, P., et al. (2024) · PubMed · Cited by 4 · see citing works
Self-citationDiCoM -- Diverse Concept Modeling towards Enhancing Generalizability in Chest X-Ray Studies
Parida, A., Capellan-Martin, D., Atito, S., et al. (2024) · arXiv · doi:10.48550/arxiv.2402.15534
Self-citation
Flagged first, then by citations
Cited by
Sorted by citations
Nanni, L., Bonato, B., & Bertoldo, A. (2025) · Sensors · doi:10.3390/s25061838 · Cited by 47 · see citing works
Liu, X., Jiang, Z., Roth, H. R., Anwar, S. M., Bonner, E. R., Mahtabfar, A., Packer, R. J., Kazerooni, A. F., Bornhorst, M., & Linguraru, M. G. (2024) · Neuro-Oncology Advances · doi:10.1093/noajnl/vdae108 · Cited by 12 · see citing works
“26 Our results are also comparable with results of the winning method (TC Dice = 0.78, WT Dice = 0.82) 40 for the pediatric BraTS challenge 2023.”
InfluentialSelf-citationA New Logic for Pediatric Brain Tumor Segmentation
Bengtsson, M., Keleş, E., Durak, G., Anwar, S. M., Velichko, Y., Linguraru, M. G., Waanders, A. J., & Bağcı, U. (2025) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi60581.2025.10980809 · Cited by 7 · see citing works
“When comparing 3L/WT models on PED BraTS 2024 testing set against baselines, the original labeling system was used without conversion to NC.”
InfluentialSelf-citationDSA : Deep Self‐Attention Medical Transformer Neuro‐Technology for Brain Tumor Segmentation
Siddiqah, M., Javed, K., Gilani, S. O., et al. (2025) · International Journal of Imaging Systems and Technology · doi:10.1002/ima.70109 · Cited by 5 · see citing works
An edge enhanced 3D mamba U‐Net for pediatric brain tumor segmentation with transfer learning
Sun, X., He, W., Ruan, J., et al. (2025) · Medical Physics · doi:10.1002/mp.70002 · Cited by 3 · see citing works
“In the ET region,Capellán et al.’s method 33 reached the highest Dice (0.733) through ensemble learning and extensive post-processing, ranking first in the BraTS 2023 Pediatric Brain Tumor Segmentation Challenge, but the proposed method showed a much lower Hausdorff distance compared with others.”
InfluentialHybrid Local-Window-Attention–Assisted U-Net Model for Multimodal Medical-Image Segmentation
Kim, J., Jin, S., Noh, Y., Moon, H., Lee, M., & Noh, W. (2025) · Lecture notes in computer science · doi:10.1007/978-3-032-04937-7_22 · Cited by 2 · see citing works
Improving Pre-trained Adult Glioma Segmentation Models Using only Post-processing Techniques
Parida, A., Capellán-Martín, D., Jiang, Z., et al. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-16365-3_22 · Cited by 1 · see citing works
“Previous BraTS winners have shown that post-processing steps such as ensemble voting [21], size-sensitive connected component filtering [9], ET-to-NCR relabeling heuristics [9], and adaptive refinement [23,16] can significantly improve segmentation accuracy.”
Self-citationMulti-Institutional Annotated Multiparametric MRI Dataset of Pediatric High-Grade Gliomas
Kazerooni, A. F., Jiang, Z., Gandhi, D., et al. (2026) · Radiology Artificial Intelligence · doi:10.1148/ryai.250902 · Cited by 1 · see citing works
Self-citationElzayat, A. R., Hanafy, N., Magdy, M., et al. (2026) · Scientific Reports · doi:10.1038/s41598-026-35627-x
Hsu, T.-L., Nguyen, D. K., Lin, P., Lin, C., & Wang, W. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-16365-3_47
FreqDualNet: frequency-aware vision transformers for tumor segmentation
Lee, J., & Lee, M. (2026) · PeerJ Computer Science · doi:10.7717/peerj-cs.3801
A novel hybrid framework integrating GA-driven 3D ResUNetGAN for MRI brain tumor segmentation
Muthulakshmi, K., & Mohan, J. (2026) · PLoS ONE · doi:10.1371/journal.pone.0349451
Hierarchical two-stage ensemble of dilated U-net models for brain tumor segmentation
Koniukhov, V. D. (2026) · Visual Computing for Industry Biomedicine and Art · doi:10.1186/s42492-026-00230-4
Shah, K., Masanagi, R., Saboowala, M., et al. (2026) · International Journal on Smart Sensing and Intelligent Systems · doi:10.2478/ijssis-2026-0063
Al-Fakih, A., Choi, K. S., Kim, D., et al. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-38072-2_10
- Authors
- Capellán-Martín, D., Jiang, Z., Parida, A., Liu, X., Lam, V., Nisar, H., Tapp, A., Elsharkawi, S., Ledesma-Carbayo, M. J., Anwar, S. M., & Linguraru, M. G.
- Affiliations
- Children's National (US)
- Corresponding author
- Linguraru, Marius George
- ORCID
- Venue
- Lecture Notes in Computer Science
- Publisher
- Springer Nature Switzerland
- Published
- 2024
- Volume
- 14669
- Pages
- 221-232
- ISSN
- 0302-9743 (ISSN-L)
- 1611-3349
- PMC
- PMC12720496
- OpenAlex
- W4405823737
- Semantic Scholar
- a39ec32366030e15e36181c13f9e9726d6054470
- Cited by
- 15
- References
- 22
- Citations by year
- 2026: 9 · 2025: 5 · 2024: 1
- FWCI
- 6.97
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- Lecture notes in computer science (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
- PubMed Central (submittedVersion)
- arXiv (Cornell University)
- Funders
- European Commission (101100633)
- Universidad Politécnica de Madrid
- Ministerio de Ciencia e Innovación (PDC2022, PID2022)
- Agencia Estatal de Investigación (PDC2022-133865-I00)
- National Cancer Institute (UG3CA236536)
- Topics
- Life Sciences › Neuroscience › Neurology › Brain Tumor Detection and Classification
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
- Health Sciences › Medicine › Genetics › Glioma Diagnosis and Treatment
- Keywords
- brain tumor segmentation, ensemble learning, BraTS, pediatric brain tumors, intracranial meningioma, brain metastases, nnU-Net, Swin UNETR, post-processing, multiparametric MRI
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-21.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-21.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
