arxiv: 2603.24985 · v2 · submitted 2026-03-26 · 💻 cs.CV

Few-Shot Left Atrial Wall Segmentation in 3D LGE MRI via Meta-Learning

Yusri Al-Sanaani , Rebecca Thornhill , Pablo Nery , Elena Pena , Robert deKemp , Calum Redpath , David Birnie , Sreeraman Rajan This is my paper

Pith reviewed 2026-05-15 00:49 UTC · model grok-4.3

classification 💻 cs.CV

keywords few-shot segmentationmeta-learningleft atrial wall3D LGE MRIMAMLboundary-aware losscardiac imagingdomain shift robustness

0 comments

The pith

Meta-learning with auxiliary cavity tasks improves few-shot accuracy for thin left atrial wall segmentation in 3D MRI.

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

The paper tests whether a Model-Agnostic Meta-Learning framework can segment the thin left atrial wall from 3D late gadolinium enhancement MRI using only five to twenty labeled examples. It meta-trains the model on the wall task together with auxiliary left and right atrial cavity segmentation tasks and applies a boundary-aware composite loss to prioritize accurate thin boundaries. Performance is measured on a hold-out test set, under an unseen synthetic domain shift, and on a separate local cohort, with direct comparison to standard supervised fine-tuning. If the approach works, it would let clinicians obtain reliable wall measurements from new patients with far fewer expert annotations than full supervision requires.

Core claim

The MAML framework meta-trained on the wall task together with auxiliary left and right atrial cavity tasks and uses a boundary-aware composite loss to emphasize thin-structure accuracy. On the hold-out test set it improves segmentation over supervised fine-tuning, reaching a Dice score of 0.64 versus 0.52 and an HD95 of 5.70 mm versus 7.60 mm at five shots, and approaches the fully supervised reference of 0.71 Dice at twenty shots. The gains persist, though reduced, under synthetic shift and on the local cohort.

What carries the argument

Model-Agnostic Meta-Learning (MAML) meta-trained jointly on the thin-wall segmentation task plus auxiliary left and right atrial cavity tasks, guided by a boundary-aware composite loss.

If this is right

At five shots the meta-trained model reaches 0.64 Dice and 5.70 mm HD95 on hold-out data versus 0.52 Dice and 7.60 mm for fine-tuning.
Under an unseen synthetic domain shift the same five-shot model still attains 0.59 Dice and 5.99 mm HD95.
Performance on a distinct local cohort remains competitive at 0.57 Dice for five shots and rises with more shots.
At twenty shots the meta-trained model reaches 0.69 Dice, nearly matching the fully supervised reference of 0.71.

Where Pith is reading between the lines

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

The same joint meta-training pattern could be tested on other thin-walled structures such as vessel walls or cortical layers where annotations are costly.
If the auxiliary-task benefit holds, the labeling burden for new clinical sites could drop from hundreds of volumes to a handful, speeding routine use of 3D atrial remodeling assessment.
The boundary-aware loss term may be worth isolating in future ablations to check whether it alone explains most of the thin-structure gains.

Load-bearing premise

The auxiliary cavity segmentation tasks share enough structure with the thin-wall task that meta-training transfers useful knowledge to new data.

What would settle it

Running the identical MAML setup without the auxiliary cavity tasks on the five-shot hold-out test set and finding no Dice or HD95 improvement over standard fine-tuning would falsify the central claim.

Figures

Figures reproduced from arXiv: 2603.24985 by Calum Redpath, David Birnie, Elena Pena, Pablo Nery, Rebecca Thornhill, Robert deKemp, Sreeraman Rajan, Yusri Al-Sanaani.

**Figure 1.** Figure 1: Meta-learning pipline for left atrium segemation. The outer loop then updates the initialization 𝜃 and the representation parameters ϕ to improve post-adaptation performance across all tasks in the batch: ℒmeta(𝜃,𝜙) = ∑ℒ𝜏𝑖 𝑁 𝑖=1 (𝑄𝜏𝑖 ; 𝜃𝑖 ′ ,𝜙), (2) (𝜃,𝜙) ← (𝜃,𝜙)− 𝛽∇(𝜃,𝜙)ℒmeta(𝜃,𝜙), (3) with meta learning rate 𝛽. Second-order MAML is memory-intensive in 3D because it backpropagates through the inner-loop u… view at source ↗

**Figure 2.** Figure 2: Axial mid-slice comparison of LA wall segmentation with 𝐾 = 5. Cases are drawn from (a) unseen domain 𝑑0 subjects and (b) the local cohort 𝑑ext. On the clean domain 𝑑0 (Table I), MAML exhibits a smaller episode-to-episode spread than FT, especially at K=5, suggesting reduced sensitivity to which subjects are selected for the support set. This is impactful in a deployment-style 𝐾-shot protocol, where perfor… view at source ↗

read the original abstract

Segmenting the left atrial wall from late gadolinium enhancement magnetic resonance images (MRI) is challenging due to the wall's thin geometry, low contrast, and the scarcity of expert annotations. We propose a Model-Agnostic Meta-Learning (MAML) framework for K-shot (K = 5, 10, 20) 3D left atrial wall segmentation that is meta-trained on the wall task together with auxiliary left atrial and right atrial cavity tasks and uses a boundary-aware composite loss to emphasize thin-structure accuracy. We evaluated MAML segmentation performance on a hold-out test set and assessed robustness under an unseen synthetic shift and on a distinct local cohort. On the hold-out test set, MAML appeared to improve segmentation performance compared to the supervised fine-tuning model, achieving a Dice score (DSC) of 0.64 vs. 0.52 and HD95 of 5.70 vs. 7.60 mm at 5-shot, and approached the fully supervised reference at 20-shot (0.69 vs. 0.71 DSC). Under unseen shift, performance degraded but remained robust: at 5-shot, MAML attained 0.59 DSC and 5.99 mm HD95 on the unseen domain shift and 0.57 DSC and 6.01 mm HD95 on the local cohort, with consistent gains as K increased. These results suggest that more accurate and reliable thin-wall boundaries are achievable in low-shot adaptation, potentially enabling clinical translation with minimal additional labeling for the assessment of atrial remodeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MAML with auxiliary cavity tasks improves few-shot LA wall segmentation modestly, but ablations are needed to confirm the meta-learning benefit.

read the letter

The paper applies Model-Agnostic Meta-Learning to the problem of segmenting the thin left atrial wall in 3D late gadolinium enhancement MRI using only a handful of labeled examples. They meta-train on the target wall task along with auxiliary segmentation of the left and right atrial cavities, and they add a boundary-aware term to the loss to focus on the thin structures. This combination produces better results than standard supervised fine-tuning in the low-shot regime. On the hold-out test set, the 5-shot Dice score rises from 0.52 to 0.64 and the 95th percentile Hausdorff distance drops from 7.60 mm to 5.70 mm. Performance improves further with more shots and stays usable under a synthetic domain shift and on an independent local cohort. The work is new in its specific tailoring of MAML plus auxiliary cavity tasks to this exact clinical segmentation challenge. The evaluation across multiple test conditions is a plus. The main limitation is the lack of ablations that would show whether the auxiliary cavity tasks contribute beyond the boundary loss itself. The cavities are much thicker and higher contrast than the atrial wall, so it is not obvious that they provide useful inductive bias for the thin-wall problem. The abstract gives no statistical significance tests and omits the full training protocol, so the quantitative claims rest on moderate evidence. This paper is aimed at researchers who need practical few-shot methods for cardiac MRI analysis. It would be worth a serious referee's time because the task is clinically relevant and the approach is a reasonable extension of existing meta-learning techniques, even though additional controls would make the contribution clearer.

Referee Report

2 major / 1 minor

Summary. The paper claims that a MAML framework meta-trained jointly on the left atrial wall segmentation task plus auxiliary left and right atrial cavity tasks, combined with a boundary-aware composite loss, enables effective K-shot (K=5,10,20) 3D wall segmentation in LGE MRI. It reports improved performance over supervised fine-tuning baselines (DSC 0.64 vs. 0.52 and HD95 5.70 vs. 7.60 mm at 5-shot on hold-out data) that approaches fully supervised results at 20-shot, with maintained robustness under synthetic domain shift and on a local cohort.

Significance. If the meta-transfer from auxiliary cavity tasks to thin-wall segmentation is confirmed, the work would be significant for few-shot medical image segmentation by reducing annotation burden for thin, low-contrast structures relevant to atrial remodeling assessment. The multi-cohort evaluation (hold-out, synthetic shift, local) is a strength that supports robustness claims. Credit is given for the explicit K-shot protocol and boundary-aware loss design.

major comments (2)

[Methods] Methods section: the central claim that auxiliary cavity tasks provide useful inductive bias for meta-transfer to the thin-wall task is load-bearing, yet no ablation is described that trains MAML on the wall task alone (without cavities) or compares against a non-meta model using only the boundary-aware loss; without this, the DSC gain (0.64 vs. 0.52) cannot be attributed to MAML rather than the loss.
[Results] Results section: reported quantitative improvements (DSC, HD95) lack statistical tests, standard deviations across runs, or confidence intervals, so it is unclear whether the observed gains over supervised fine-tuning are statistically reliable given the small K-shot regimes and test-set sizes.

minor comments (1)

[Abstract] Abstract: the statement that performance 'approached the fully supervised reference at 20-shot (0.69 vs. 0.71 DSC)' would benefit from explicitly stating the fully supervised DSC value and whether the difference is within expected variance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important aspects of our experimental design and reporting. We address each major comment below and will update the manuscript to strengthen the claims regarding the contribution of auxiliary tasks and to include appropriate statistical analysis.

read point-by-point responses

Referee: [Methods] Methods section: the central claim that auxiliary cavity tasks provide useful inductive bias for meta-transfer to the thin-wall task is load-bearing, yet no ablation is described that trains MAML on the wall task alone (without cavities) or compares against a non-meta model using only the boundary-aware loss; without this, the DSC gain (0.64 vs. 0.52) cannot be attributed to MAML rather than the loss.

Authors: We agree that the current manuscript lacks the necessary ablations to isolate the effect of the auxiliary cavity tasks within the MAML framework and to separate the contribution of meta-learning from the boundary-aware loss. In the revised version, we will add an ablation study that includes: (1) MAML trained solely on the wall segmentation task without auxiliary cavities, and (2) a non-meta supervised baseline using only the boundary-aware composite loss. These results will be reported alongside the main experiments to clarify the source of the observed performance gains. revision: yes
Referee: [Results] Results section: reported quantitative improvements (DSC, HD95) lack statistical tests, standard deviations across runs, or confidence intervals, so it is unclear whether the observed gains over supervised fine-tuning are statistically reliable given the small K-shot regimes and test-set sizes.

Authors: We acknowledge that the reported metrics would benefit from statistical rigor to confirm reliability, especially in low-shot settings. In the revision, we will rerun the experiments across multiple random seeds for K-shot task sampling, report mean and standard deviation for DSC and HD95, and include paired statistical tests (e.g., Wilcoxon signed-rank test) with p-values to assess significance of differences versus the fine-tuning baseline. Confidence intervals will also be added where feasible. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results measured on independent test sets

full rationale

The paper reports an empirical MAML-based segmentation method evaluated via direct measurement of DSC and HD95 on hold-out test sets, an unseen synthetic shift, and a distinct local cohort. Performance numbers are obtained from model inference on separate data rather than derived from any fitted parameter or self-referential equation. No self-definitional loops, fitted-input predictions, or load-bearing self-citations appear in the derivation; the auxiliary-task meta-training and boundary-aware loss are standard components whose benefit is assessed by ablation-free but externally measured comparison to supervised fine-tuning. The chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the standard MAML assumption that meta-training across related tasks enables fast adaptation; no new free parameters are introduced beyond conventional meta-learning hyperparameters, and no new entities are postulated.

free parameters (1)

K (number of shots)
Chosen values 5, 10, 20 for evaluation; not fitted but selected to demonstrate low-data regime.

axioms (1)

domain assumption Auxiliary left and right atrial cavity segmentation tasks share sufficient structure with the thin-wall task to support positive meta-transfer.
Invoked in the meta-training design described in the abstract.

pith-pipeline@v0.9.0 · 5613 in / 1316 out tokens · 29781 ms · 2026-05-15T00:49:00.820846+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

[1]

Recent Advances in Fibrosis and Scar Segmentation From Cardiac MRI: A State-of-the-Art Review and Future Perspectives,

Y. Wu, Z. Tang, B. Li, D. Firmin, and G. Yang, “Recent Advances in Fibrosis and Scar Segmentation From Cardiac MRI: A State-of-the-Art Review and Future Perspectives,” Front. Physiol., vol. 12, p. 709230, Aug. 2021, doi: 10.3389/FPHYS.2021.709230

work page doi:10.3389/fphys.2021.709230 2021
[2]

RAS Dataset: A 3D Cardiac LGE -MRI Dataset for Segmentation of Right Atrial Cavity,

J. Zhu et al. , “RAS Dataset: A 3D Cardiac LGE -MRI Dataset for Segmentation of Right Atrial Cavity,” Sci. Data, vol. 11, no. 1, p. 401, Dec. 2024, doi: 10.1038/S41597-024-03253-9

work page doi:10.1038/s41597-024-03253-9 2024
[3]

The use of MRI in quantification of the atrial fibrosis in patients with rheumatic mitral disease,

A. S. Ismail, Y. Baghdady, M. A. Salem, and A. A. Wahab, “The use of MRI in quantification of the atrial fibrosis in patients with rheumatic mitral disease,” Egyptian Journal of Radiology and Nuclear Medicine, vol. 51, no. 1, Dec. 2020, doi: 10.1186/s43055-020-00322-y

work page doi:10.1186/s43055-020-00322-y 2020
[4]

Left atrial fibrosis in atrial fibrillation: Mechanisms, clinical evaluation and management,

J. Ma, Q. Chen, and S. Ma, “Left atrial fibrosis in atrial fibrillation: Mechanisms, clinical evaluation and management,” Mar. 01, 2021, John Wiley and Sons Inc. doi: 10.1111/jcmm.16350

work page doi:10.1111/jcmm.16350 2021
[5]

Medical image analysis on left atrial LGE MRI for atrial fibrillation studies: A review,

L. Li, V. A. Zimmer, J. A. Schnabel, and X. Zhuang, “Medical Image Analysis on Left Atrial LGE MRI for Atrial Fibrillation Studies: A Review,” Med. Image Anal. , vol. 77, p. 102360, Apr. 2022, doi: 10.1016/J.MEDIA.2022.102360

work page doi:10.1016/j.media.2022.102360 2022
[6]

Usformer: A small network for left atrium segmentation of 3D LGE MRI,

H. Lin et al., “Usformer: A small network for left atrium segmentation of 3D LGE MRI,” Heliyon, vol. 10, no. 7, Apr. 2024, doi: 10.1016/j.heliyon.2024.e28539

work page doi:10.1016/j.heliyon.2024.e28539 2024
[7]

A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging,

Z. Xiong et al., “A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging,” Med. Image Anal. , vol. 67, Jan. 2021, doi: 10.1016/j.media.2020.101832

work page doi:10.1016/j.media.2020.101832 2021
[8]

Simultaneous left atrium anatomy and scar segmentations via deep learning in multiview information with attention,

G. Yang et al. , “Simultaneous left atrium anatomy and scar segmentations via deep learning in multiview information with attention,” Future Generation Computer Systems , vol. 107, pp. 215 – 228, Jun. 2020, doi: 10.1016/j.future.2020.02.005

work page doi:10.1016/j.future.2020.02.005 2020
[9]

AtrialJSQnet: A New framework for joint segmentation and quantification of left atrium and scars incorporating spatial and shape information,

L. Li, V. A. Zimmer, J. A. Schnabel, and X. Zhuang, “AtrialJSQnet: A New framework for joint segmentation and quantification of left atrium and scars incorporating spatial and shape information,” Med. Image Anal., vol. 76, Feb. 2022, doi: 10.1016/j.media.2021.102303

work page doi:10.1016/j.media.2021.102303 2022
[10]

Evaluating Convolution, Attention, and Mamba Based U -Net Models for Multi- class Bi -Atrial Segmentation from LGE-MRI,

C. Thesing, A. Bueno -Orovio, and A. Banerjee, “Evaluating Convolution, Attention, and Mamba Based U -Net Models for Multi- class Bi -Atrial Segmentation from LGE-MRI,” in Lecture Notes in Computer Science, Springer Science and Business Media Deutschland GmbH, 2025, pp. 214–225. doi: 10.1007/978-3-031-87756-8_22

work page doi:10.1007/978-3-031-87756-8_22 2025
[11]

Multi-loss 3D Segmentation for Enhanced Bi- atrial Segmentation,

E. Almar-Munoz et al., “Multi-loss 3D Segmentation for Enhanced Bi- atrial Segmentation,” in Lecture Notes in Computer Science, Springer Science and Business Media Deutschland GmbH, 2025, pp. 236 –244. doi: 10.1007/978-3-031-87756-8_24

work page doi:10.1007/978-3-031-87756-8_24 2025
[12]

An Ensemble of 3D Residual Encoder UNet Models for Solving Multi - class Bi-atrial Segmentation Challenge,

A. Zolotarev, K. Johnson, A. Khan, G. Slabaugh, and C. Roney, “An Ensemble of 3D Residual Encoder UNet Models for Solving Multi - class Bi-atrial Segmentation Challenge,” in Lecture Notes in Computer Science, Springer Science and Business Media Deutschland GmbH, 2025, pp. 209–213. doi: 10.1007/978-3-031-87756-8_21

work page doi:10.1007/978-3-031-87756-8_21 2025
[13]

Few-Shot Learning for Medical Image Segmentation Using 3D U-Net and Model -Agnostic Meta -Learning (MAML),

A. M. Alsaleh, E. Albalawi, A. Algosaibi, S. S. Albakheet, and S. B. Khan, “Few-Shot Learning for Medical Image Segmentation Using 3D U-Net and Model -Agnostic Meta -Learning (MAML),” Diagnostics (Basel)., vol. 14, no. 12, Jun. 2024, doi: 10.3390/DIAGNOSTICS14121213

work page doi:10.3390/diagnostics14121213 2024
[14]

Meta-learning with implicit gradients in a few -shot setting for medical image segmentation,

R. Khadka et al., “Meta-learning with implicit gradients in a few -shot setting for medical image segmentation,” Comput. Biol. Med. , vol. 143, p. 105227, Apr. 2022, doi: 10.1016/J.COMPBIOMED.2022.105227

work page doi:10.1016/j.compbiomed.2022.105227 2022
[15]

Few -shot Medical Image Segmentation with High- Fidelity Prototypes,

S. Tang et al. , “Few -shot Medical Image Segmentation with High- Fidelity Prototypes,” Med. Image Anal. , vol. 100, p. 103412, Feb. 2025, doi: 10.1016/J.MEDIA.2024.103412

work page doi:10.1016/j.media.2024.103412 2025
[16]

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks,

C. Finn, P. Abbeel, and S. Levine, “Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks,” Jul. 17, 2017, PMLR. Accessed: Jan. 13, 2026. [Online]. Available: https://proceedings.mlr.press/v70/finn17a.html

work page 2017
[17]

Learning to Segment Medical Images from Few -Shot Sparse Labels,

P. H. T. Gama, H. Oliveira, and J. A. Dos Santos, “Learning to Segment Medical Images from Few -Shot Sparse Labels,” SIBGRAPI Conference on Graphics, Patterns and Images , pp. 89 –96, 2021, doi: 10.1109/SIBGRAPI54419.2021.00021

work page doi:10.1109/sibgrapi54419.2021.00021 2021
[18]

Adaptive dynamic inference for few -shot left atrium segmentation,

J. Chen et al., “Adaptive dynamic inference for few -shot left atrium segmentation,” Med. Image Anal. , vol. 98, Dec. 2024, doi: 10.1016/j.media.2024.103321

work page doi:10.1016/j.media.2024.103321 2024
[19]

Domain Adaptation for Medical Image Analysis: A Survey,

H. Guan and M. Liu, “Domain Adaptation for Medical Image Analysis: A Survey,” IEEE Trans. Biomed. Eng. , vol. 69, no. 3, p. 1173, Mar. 2022, doi: 10.1109/TBME.2021.3117407

work page doi:10.1109/tbme.2021.3117407 2022
[20]

Domain Generalization for Medical Image Analysis: A Review,

J. S. Yoon, K. Oh, Y. Shin, M. A. Mazurowski, and H. Il Suk, “Domain Generalization for Medical Image Analysis: A Review,” Proceedings of the IEEE , vol. 112, no. 10, pp. 1583 –1609, 2024, doi: 10.1109/JPROC.2024.3507831

work page doi:10.1109/jproc.2024.3507831 2024
[21]

A Survey on Domain Generalization for Medical Image Analysis,

Z. Niu, S. Ouyang, S. Xie, Y. Chen, and L. Lin, “A Survey on Domain Generalization for Medical Image Analysis,” Feb. 2024, Accessed: Jan. 13, 2026. [Online]. Available: https://arxiv.org/pdf/2402.05035v1

work page arXiv 2024
[22]

FSDA -DG: Improving Cross-Domain Generalizability of Medical Image Segmentation with Few Source Domain Annotations ,

Z. Ye, K. Wang, W. Lv, Q. Feng, and L. Lu, “FSDA -DG: Improving Cross-Domain Generalizability of Medical Image Segmentation with Few Source Domain Annotations ,” Med. Image Anal. , vol. 105, p. 103704, Oct. 2025, doi: 10.1016/J.MEDIA.2025.103704

work page doi:10.1016/j.media.2025.103704 2025
[23]

, author Raghu, M

A. Raghu, M. Raghu, S. Bengio, and O. Vinyals, “Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML,” 8th International Conference on Learning Representations, ICLR 2020, Sep. 2019, Accessed: Jan. 22, 2026. [Online]. Available: https://arxiv.org/pdf/1909.09157

work page arXiv 2020
[24]

On First-Order Meta-Learning Algorithms,

A. Nichol, J. Achiam, and J. Schulman, “On First-Order Meta-Learning Algorithms,” arXiv.org, 2018

work page 2018
[25]

& Brox, T

O. Ronneberger, P. Fischer, and T. Brox, “U -Net: Convolutional Networks for Biomedical Image Segmentation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 9351, pp. 234– 241, 2015, doi: 10.1007/978-3-319-24574-4_28

work page doi:10.1007/978-3-319-24574-4_28 2015
[26]

3D U -net: Learning dense volumetric segmentation from sparse annotation,

Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U -net: Learning dense volumetric segmentation from sparse annotation,” Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 9901 LNCS, pp. 424–432, 2016, doi: 10.1007/978-3-319-46723-8_49/TABLES/3

work page doi:10.1007/978-3-319-46723-8_49/tables/3 2016
[27]

Boundary loss for highly unbalanced segmentation,

H. Kervadec, J. Bouchtiba, C. Desrosiers, E. Granger, J. Dolz, and I. Ben Ayed, “Boundary loss for highly unbalanced segmentation,” Med. Image Anal., vol. 67, Jan. 2021, doi: 10.1016/j.media.2020.101851

work page doi:10.1016/j.media.2020.101851 2021
[28]

& Maier-Hein, K

F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier - Hein, “nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation,” Nature Methods 2020 18:2, vol. 18, no. 2, pp. 203–211, Dec. 2020, doi: 10.1038/s41592-020-01008-z

work page doi:10.1038/s41592-020-01008-z 2020
[29]

Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy: Deep Learning Algorithm Development and Validation Study,

S. Nikolov et al. , “Clinically Applicable Segmentation of Head and Neck Anatomy for Radiotherapy: Deep Learning Algorithm Development and Validation Study,” J. Med. Internet Res., vol. 23, no. 7, Jul. 2021, doi: 10.2196/26151

work page doi:10.2196/26151 2021