arxiv: 2604.13409 · v1 · submitted 2026-04-15 · 💻 cs.CV

Recognition: unknown

CausalDisenSeg: A Causality-Guided Disentanglement Framework with Counterfactual Reasoning for Robust Brain Tumor Segmentation Under Missing Modalities

Bo Liu , Yulong Zou , Jin Hong

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:33 UTC · model grok-4.3

classification 💻 cs.CV

keywords brain tumor segmentationmissing modalitiescausal disentanglementcounterfactual reasoningstructural causal modelMRIrobustnessdeep learning

0 comments

The pith

CausalDisenSeg isolates anatomical causal factors from stylistic biases via a structural causal model and counterfactual reasoning to maintain accurate brain tumor segmentation despite missing MRI modalities.

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

The paper claims that current multimodal segmentation models exploit spurious correlations from complete MRI training data, leading to failure when modalities are absent in practice. It reframes the problem around a Structural Causal Model that explicitly separates the true anatomical causal factor from modality-specific stylistic bias. The proposed three-stage intervention uses a conditional variational autoencoder with an HSIC constraint for statistical independence, a Region Causality Module to anchor features to physical tumor locations, and dual-adversarial counterfactual reasoning to suppress the direct effect of bias. If the separation holds, the model should produce consistent segmentations from any available subset of modalities without retraining. This addresses a common clinical reality where full four-modality scans are frequently unavailable.

Core claim

CausalDisenSeg reframes missing-modality segmentation as isolating the anatomical Causal Factor from the stylistic Bias Factor through explicit causal disentanglement with CVAE and HSIC orthogonality, reinforcement via a Region Causality Module that grounds features in physical regions, and dual-adversarial counterfactual reasoning that suppresses the natural direct effect of bias while preserving the causal path, yielding higher accuracy and consistency than prior fusion methods on BraTS 2020 under severe missing-modality conditions and a macro-average DSC of 84.49 on cross-dataset BraTS 2023 evaluation.

What carries the argument

The Structural Causal Model that decomposes inputs into an anatomical Causal Factor and a stylistic Bias Factor, with disentanglement enforced by Conditional Variational Autoencoder plus HSIC constraint, feature grounding by the Region Causality Module, and bias suppression by dual-adversarial counterfactual reasoning.

If this is right

Segmentation accuracy remains high and consistent across all possible missing-modality combinations without requiring separate models for each scenario.
The approach reduces overfitting to modality-specific artifacts that appear only in complete training scans.
Cross-dataset generalization improves because the causal path focuses on anatomy rather than dataset-specific style distributions.
Clinical deployment becomes feasible in settings where scanners or protocols routinely omit one or more MRI sequences.

Where Pith is reading between the lines

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

The same causal separation principle could be tested on other multimodal tasks such as cardiac MRI or abdominal CT where data incompleteness is common.
If the dual-adversarial counterfactual step proves stable, it might serve as a general template for removing shortcut learning in any vision model trained on correlated but non-causal inputs.
Applying the framework to longitudinal scans could help isolate true tumor progression signals from changes in acquisition style over time.

Load-bearing premise

The anatomical causal factor can be isolated from the stylistic bias factor using a conditional variational autoencoder with an HSIC constraint, the Region Causality Module can ground features in physical tumor regions, and dual-adversarial counterfactual reasoning can suppress the natural direct effect of bias without harming the causal path to the segmentation output.

What would settle it

If CausalDisenSeg shows no statistically significant improvement in Dice scores compared with standard fusion baselines when evaluated on BraTS 2020 with random single-modality or dual-modality dropout, the claim that the causal intervention removes dependency on missing-modality shortcuts would be falsified.

read the original abstract

In clinical practice, the robustness of deep learning models for multimodal brain tumor segmentation is severely compromised by incomplete MRI data. This vulnerability stems primarily from modality bias, where models exploit spurious correlations as shortcuts rather than learning true anatomical structures. Existing feature fusion methods fail to fundamentally eliminate this dependency. To address this, we propose CausalDisenSeg, a novel Structural Causal Model (SCM)-grounded framework that achieves robust segmentation via causality-guided disentanglement and counterfactual reasoning. We reframe the problem as isolating the anatomical Causal Factor from the stylistic Bias Factor. Our framework implements a three-stage causal intervention: (1) Explicit Causal Disentanglement: A Conditional Variational Autoencoder (CVAE) coupled with an HSIC constraint mathematically enforces statistical orthogonality between anatomical and style features. (2) Causal Representation Reinforcement: A Region Causality Module (RCM) explicitly grounds causal features in physical tumor regions. (3) Counterfactual Reasoning: A dual-adversarial strategy actively suppresses the residual Natural Direct Effect (NDE) of the bias, forcing its spatial attention to be mutually exclusive from the causal path. Extensive experiments on the BraTS 2020 dataset demonstrate that CausalDisenSeg significantly outperforms state-of-the-art methods in accuracy and consistency across severe missing-modality scenarios. Furthermore, cross-dataset evaluation on BraTS 2023 under the same protocol yields a state-of-the-art macro-average DSC of 84.49.

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

CausalDisenSeg proposes a three-stage causal disentanglement pipeline for missing-modality brain tumor segmentation, but the abstract gives no equations, ablations, or verification that the claimed factor separation actually occurs.

read the letter

The main thing to know is that this paper takes the practical problem of missing MRI modalities in brain tumor segmentation and tries to solve it with an explicit causal intervention: disentangle anatomical factors from style biases, ground them in regions, then suppress residual bias via counterfactuals. The reported gains on BraTS 2020 across missing-modality cases and the 84.49 macro DSC on BraTS 2023 are the concrete numbers they put forward. That combination of CVAE plus HSIC, a Region Causality Module, and dual-adversarial NDE suppression is the specific new piece they claim over prior disentanglement work. The framing around modality bias as a shortcut that needs causal removal is clear and matches a real deployment issue. The cross-dataset check adds a bit of weight to the empirical side. The soft spots are exactly where the stress-test note points. The abstract states the stages but shows none of the loss equations, no ablation tables, no statistical significance tests, and no post-hoc checks on whether the learned factors stay invariant under modality dropout or whether the attention maps are actually mutually exclusive. Without those, it is impossible to tell whether the outperformance comes from the causal mechanism or from the extra capacity and regularization. The assumption that CVAE+HSIC plus adversarial training cleanly isolates the anatomical causal factor from stylistic bias is load-bearing and untested in what is visible. If the full paper contains the derivations and those checks, the picture changes; on the abstract alone the soundness is low. This paper is for researchers working on robust multimodal medical segmentation who want to see causal ideas applied to the missing-modality setting. A reader already familiar with CVAE disentanglement and HSIC would get the most out of the architecture description and the specific three-stage design. It deserves a serious referee because the clinical problem is important, the combination is new enough to be worth evaluating, and the empirical numbers are stated clearly enough to be checked. I would send it to review with a request for the missing methods details and verification experiments rather than desk-reject.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes CausalDisenSeg, a Structural Causal Model (SCM)-grounded framework for robust brain tumor segmentation under missing MRI modalities. It reframes the problem as disentangling anatomical Causal Factors from stylistic Bias Factors using a three-stage intervention: (1) CVAE with HSIC constraint for explicit causal disentanglement, (2) Region Causality Module (RCM) for grounding features in physical regions, and (3) dual-adversarial counterfactual reasoning to suppress the Natural Direct Effect (NDE) of bias. Experiments on BraTS 2020 demonstrate significant outperformance in severe missing-modality scenarios, and cross-dataset evaluation on BraTS 2023 achieves a state-of-the-art macro-average DSC of 84.49.

Significance. Should the causal disentanglement and counterfactual mechanisms prove effective in isolating true anatomical structures independent of modality biases, this could offer a substantial improvement over existing feature fusion methods in handling incomplete multimodal data, a common issue in clinical settings. The framework's emphasis on causality provides a principled approach that may generalize better. The inclusion of cross-dataset evaluation on BraTS 2023 is a strength, supporting broader applicability. However, the ultimate significance depends on rigorous evidence that the performance gains arise from the proposed causal interventions rather than incidental regularization effects.

major comments (3)

Abstract and Methods (Stage 1): The claim that the CVAE coupled with HSIC constraint 'mathematically enforces statistical orthogonality' between anatomical and style features is central to the disentanglement stage, but the abstract and provided description contain no explicit loss equations, HSIC formulation, or derivation showing how this holds under missing modalities. This undermines verification of whether the intervention achieves causal separation or merely reduces correlations.
Experiments: The reported outperformance on BraTS 2020 lacks details on the specific missing-modality scenarios tested, ablation studies isolating the contribution of each stage (disentanglement, RCM, counterfactual), and statistical significance tests, which are necessary to support that the gains stem from the causal intervention rather than standard regularization.
Counterfactual Reasoning Stage: The dual-adversarial strategy to suppress residual NDE while preserving the causal path is load-bearing for the robustness claims, yet no post-hoc checks (e.g., whether attention maps are mutually exclusive or factors invariant to modality dropout) are mentioned. This leaves the risk that the adversarial objective may leak causal information, explaining results via ordinary training dynamics.

minor comments (2)

The acronym 'NDE' for Natural Direct Effect should be introduced with a brief definition or reference to causal inference literature upon first use.
Ensure citations to foundational works on HSIC for independence and counterfactual reasoning in vision tasks are included to contextualize the novelty.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. These have prompted us to enhance the clarity of the mathematical formulations, experimental reporting, and validation analyses. We address each major comment point by point below, indicating the revisions made to the manuscript.

read point-by-point responses

Referee: Abstract and Methods (Stage 1): The claim that the CVAE coupled with HSIC constraint 'mathematically enforces statistical orthogonality' between anatomical and style features is central to the disentanglement stage, but the abstract and provided description contain no explicit loss equations, HSIC formulation, or derivation showing how this holds under missing modalities. This undermines verification of whether the intervention achieves causal separation or merely reduces correlations.

Authors: We appreciate the referee highlighting the need for explicit mathematical details to verify the causal separation. The abstract is intentionally concise and omits equations per standard practice, but the Methods section (specifically the formulation of Stage 1) provides the full CVAE objective and HSIC constraint. The HSIC term is defined as HSIC(Z_a, Z_s) = ||C_{Z_a Z_s}||^2_{HS}, where Z_a and Z_s denote the anatomical causal and stylistic bias latents, respectively, using the Hilbert-Schmidt norm of the cross-covariance operator in RKHS. This is minimized jointly with the CVAE ELBO, and the conditioning on available modalities in the CVAE encoder ensures the orthogonality holds even under missing data by preventing modality-specific shortcuts from leaking into Z_a. We have now inserted the explicit loss equations, HSIC definition, and a short derivation of independence under partial observations into the revised Methods section for direct verification. revision: yes
Referee: Experiments: The reported outperformance on BraTS 2020 lacks details on the specific missing-modality scenarios tested, ablation studies isolating the contribution of each stage (disentanglement, RCM, counterfactual), and statistical significance tests, which are necessary to support that the gains stem from the causal intervention rather than standard regularization.

Authors: We agree that granular experimental details are essential to attribute gains specifically to the causal components. The Experiments section details the missing-modality protocols on BraTS 2020, encompassing all combinations of 1-, 2-, and 3-modality dropouts with random and fixed patterns. Ablation studies isolating each stage (CVAE+HSIC disentanglement, RCM, and dual-adversarial counterfactual) are reported in Table 3, with incremental DSC improvements shown. To further substantiate that improvements arise from causal interventions rather than generic regularization, we have added paired statistical significance tests (Wilcoxon signed-rank tests with p < 0.05) comparing against baselines and ablated variants in the revised results. These additions are now explicitly referenced in the main text and supplementary material. revision: yes
Referee: Counterfactual Reasoning Stage: The dual-adversarial strategy to suppress residual NDE while preserving the causal path is load-bearing for the robustness claims, yet no post-hoc checks (e.g., whether attention maps are mutually exclusive or factors invariant to modality dropout) are mentioned. This leaves the risk that the adversarial objective may leak causal information, explaining results via ordinary training dynamics.

Authors: This concern about potential leakage in the dual-adversarial counterfactual stage is well-taken and directly tests the mechanism's effectiveness. In the revised manuscript, we have incorporated post-hoc validation analyses for this stage. These include qualitative visualizations of spatial attention maps confirming mutual exclusivity between the causal anatomical path and the bias path, as well as quantitative invariance tests measuring feature stability (via cosine similarity) of the causal latents under varying modality dropout rates. The results demonstrate that the adversarial suppression of NDE does not leak causal information into the bias pathway, distinguishing the approach from standard adversarial regularization. These checks are presented in a new subsection of the Experiments with corresponding figures. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical architecture with independent validation

full rationale

The paper describes a three-stage framework (CVAE+HSIC disentanglement, Region Causality Module, dual-adversarial counterfactual suppression) whose performance claims rest on experimental results on BraTS 2020/2023 rather than any closed-form derivation. No equations reduce outputs to inputs by construction, no fitted parameters are relabeled as predictions, and no load-bearing steps depend on self-citations or imported uniqueness theorems. The SCM framing is used as motivation for the architecture, not as a mathematical reduction that collapses to the method's own assumptions.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

The framework rests on the assumption that modality bias can be modeled as a separable stylistic factor whose natural direct effect can be adversarially removed; several modeling choices (HSIC weight, adversarial balance) function as free parameters whose values are not reported.

free parameters (2)

HSIC constraint weight
Strength of the independence penalty between anatomical and style features in the CVAE is a tunable hyperparameter required to enforce orthogonality.
Adversarial loss coefficients
Weights balancing the dual-adversarial suppression of bias NDE are chosen to achieve mutual exclusivity of attention maps.

axioms (2)

domain assumption The segmentation problem can be reframed as isolating an anatomical Causal Factor from a stylistic Bias Factor
Explicitly stated as the core reframing of the problem in the abstract.
domain assumption CVAE plus HSIC constraint mathematically enforces statistical orthogonality between anatomical and style features
Assumed to hold as the mechanism for explicit causal disentanglement.

invented entities (2)

Region Causality Module (RCM) no independent evidence
purpose: Explicitly grounds causal features in physical tumor regions
New module introduced to reinforce causal representation.
Natural Direct Effect (NDE) of the bias no independent evidence
purpose: Models the direct influence of stylistic bias that must be suppressed via counterfactual reasoning
Causal concept applied to force spatial attention exclusivity.

pith-pipeline@v0.9.0 · 5569 in / 1665 out tokens · 72806 ms · 2026-05-10T13:33:12.164558+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

36 extracted references · 1 canonical work pages

[1]

Segment anything in medical images,

J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, "Segment anything in medical images," Nature communications, vol. 15, no. 1, p. 654, 2024

2024
[2]

Brain tumor segmentation with deep neural networks,

M. Havaei et al. , "Brain tumor segmentation with deep neural networks," Medical image analysis, vol. 35, pp. 18-31, 2017

2017
[3]

Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,

H. Sung et al. , "Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries," CA: a cancer journal for clinicians, vol. 71, no. 3, pp. 209-249, 2021

2020
[4]

Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion,

C. Chen, Q. Dou, Y . Jin, H. Chen, J. Qin, and P. -A. Heng, "Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion," in International conference on medical image computing and computer-assisted intervention, 2019: Springer, pp. 447-456

2019
[5]

nnU -Net: a self - configuring method for deep learning-based biomedical image segmentation,

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier -Hein, "nnU -Net: a self - configuring method for deep learning-based biomedical image segmentation," Nature methods, vol. 18, no. 2, pp. 203-211, 2021

2021
[6]

Incomplete Scans and Lost Revenue In MRI,

M. Vibberts, "Incomplete Scans and Lost Revenue In MRI," ed, 2021

2021
[7]

A literature survey of MR-based brain tumor segmentation with missing modalities,

T. Zhou, S. Ruan, and H. Hu, "A literature survey of MR-based brain tumor segmentation with missing modalities," Computerized Medical Imaging and Graphics, vol. 104, p. 102167, 2023

2023
[8]

Missing MRI pulse sequence synthesis using multi -modal generative adversarial network,

A. Sharma and G. Hamarneh, "Missing MRI pulse sequence synthesis using multi -modal generative adversarial network," IEEE transactions on medical imaging, vol. 39, no. 4, pp. 1170-1183, 2019

2019
[9]

M3AE: multimodal representation learning for brain tumor segmentation with missing modalities,

H. Liu, D. Wei, D. Lu, J. Sun, L. Wang, and Y . Zheng, "M3AE: multimodal representation learning for brain tumor segmentation with missing modalities," in Proceedings of the AAAI conference on artificial intelligence, 2023, vol. 37, no. 2, pp. 1657-1665

2023
[10]

Hetero -modal variational encoder-decoder for joint modality completion and segmentation,

R. Dorent, S. Joutard, M. Modat, S. Ourselin, and T. Vercauteren, "Hetero -modal variational encoder-decoder for joint modality completion and segmentation," in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2019: Springer, pp. 74-82

2019
[11]

D 2-Net: Dual disentanglement network for brain tumor segmentation with missing modalities,

Q. Yang, X. Guo, Z. Chen, P. Y . Woo, and Y . Yuan, "D 2-Net: Dual disentanglement network for brain tumor segmentation with missing modalities," IEEE Transactions on Medical Imaging, vol. 41, no. 10, pp. 2953-2964, 2022

2022
[12]

RFNet: Region-aware fusion network for incomplete multi - modal brain tumor segmentation,

Y . Ding, X. Y u, and Y . Yang, "RFNet: Region-aware fusion network for incomplete multi - modal brain tumor segmentation," in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3975-3984

2021
[13]

Multi -modal learning with missing modality via shared -specific feature modelling,

H. Wang, Y . Chen, C. Ma, J. Avery, L. Hull, and G. Carneiro, "Multi -modal learning with missing modality via shared -specific feature modelling," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 15878-15887

2023
[14]

Multimodal invariant feature prompt network for brain tumor segmentation with missing modalities,

Y . Diao, H. Fang, H. Yu, F. Li, and Y . Xu, "Multimodal invariant feature prompt network for brain tumor segmentation with missing modalities," Neurocomputing, vol. 616, p. 128847, 2025

2025
[15]

Missing as masking: arbitrary cross -modal feature reconstruction for incomplete multimodal brain tumor segmentation,

Z. Zeng, Z. Peng, X. Yang, and W. Shen, "Missing as masking: arbitrary cross -modal feature reconstruction for incomplete multimodal brain tumor segmentation," in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2024: Springer, pp. 424-433

2024
[16]

Pearl, M

J. Pearl, M. Glymour, and N. P. Jewell, Causal inference in statistics: A primer. John Wiley & Sons, 2016

2016
[17]

On Causal and Anticausal Learning

B. Schölkopf, D. Janzing, J. Peters, E. Sgouritsa, K. Zhang, and J. Mooij, "On causal and anticausal learning," arXiv preprint arXiv:1206.6471, 2012

work page Pith review arXiv 2012
[18]

Deep multimodal fusion by channel exchanging,

Y . Wang, W. Huang, F. Sun, T. Xu, Y . Rong, and J. Huang, "Deep multimodal fusion by channel exchanging," Advances in neural information processing systems, vol. 33, pp. 4835-4845, 2020

2020
[19]

Adaptive mixtures of local experts,

R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, "Adaptive mixtures of local experts," Neural computation, vol. 3, no. 1, pp. 79-87, 1991

1991
[20]

Pearl and D

J. Pearl and D. Mackenzie, The book of why: the new science of cause and effect . Basic books, 2018

2018
[21]

Peters, D

J. Peters, D. Janzing, and B. Scholkopf, Elements of causal inference: foundations and learning algorithms. MIT press, 2017

2017
[22]

Analysis of representations for domain adaptation,

S. Ben-David, J. Blitzer, K. Crammer, and F. Pereira, "Analysis of representations for domain adaptation," Advances in neural information processing systems, vol. 19, 2006

2006
[23]

Domain adaptation by using causal inference to predict invariant conditional distributions,

S. Magliacane, T. Van Ommen, T. Claassen, S. Bongers, P. Versteeg, and J. M. Mooij, "Domain adaptation by using causal inference to predict invariant conditional distributions," Advances in neural information processing systems, vol. 31, 2018

2018
[24]

ACN: adversarial co -training network for brain tumor segmentation with missing modalities,

Y . Wang et al. , "ACN: adversarial co -training network for brain tumor segmentation with missing modalities," in International conference on medical image computing and computer - assisted intervention, 2021: Springer, pp. 410-420

2021
[25]

Hemis: Hetero -modal image segmentation,

M. Havaei, N. Guizard, N. Chapados, and Y . Bengio, "Hemis: Hetero -modal image segmentation," in International conference on medical image computing and computer-assisted intervention, 2016: Springer, pp. 469-477

2016
[26]

SMU -Net: Style matching U -Net for brain tumor segmentation with missing modalities,

R. Azad, N. Khosravi, and D. Merhof, "SMU -Net: Style matching U -Net for brain tumor segmentation with missing modalities," in International conference on medical imaging with deep learning, 2022: PMLR, pp. 48-62

2022
[27]

mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation,

Y . Zhang et al. , "mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation," in International conference on medical image computing and computer-assisted intervention, 2022: Springer, pp. 107-117

2022
[28]

Mftrans: Modality-masked fusion transformer for incomplete multi -modality brain tumor segmentation,

J. Shi, L. Yu, Q. Cheng, X. Yang, K.-T. Cheng, and Z. Yan, "Mftrans: Modality-masked fusion transformer for incomplete multi -modality brain tumor segmentation," IEEE journal of biomedical and health informatics, vol. 28, no. 1, pp. 379-390, 2023

2023
[29]

A2fseg: Adaptive multi -modal fusion network for medical image segmentation,

Z. Wang and Y . Hong, "A2fseg: Adaptive multi -modal fusion network for medical image segmentation," in International Conference on Medical Image Computing and Computer - Assisted Intervention, 2023: Springer, pp. 673-681

2023
[30]

Im-fuse: A mamba- based fusion block for brain tumor segmentation with incomplete modalities,

V . Pipoli, A. Saporita, K. Marchesini, C. Grana, E. Ficarra, and F. Bolelli, "Im-fuse: A mamba- based fusion block for brain tumor segmentation with incomplete modalities," in International Conference on Medical Image Computing and Computer-Assisted Intervention, 2025: Springer, pp. 225-235

2025
[31]

Does adding a modality really make a positive impacts in incomplete multi-modal brain tumor segmentation?,

Y . Qiu, K. Jiang, H. Yao, Z. Wang, and S. i. Satoh, "Does adding a modality really make a positive impacts in incomplete multi-modal brain tumor segmentation?," IEEE Transactions on Medical Imaging, 2025

2025
[32]

A kernel statistical test of independence,

A. Gretton, K. Fukumizu, C. Teo, L. Song, B. Schölkopf, and A. Smola, "A kernel statistical test of independence," Advances in neural information processing systems, vol. 20, 2007

2007
[33]

Dropout as a bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, "Dropout as a bayesian approximation: Representing model uncertainty in deep learning," in international conference on machine learning , 2016: PMLR, pp. 1050-1059

2016
[34]

The HSIC bottleneck: Deep learning without back- propagation,

W.-D. K. Ma, J. Lewis, and W. B. Kleijn, "The HSIC bottleneck: Deep learning without back- propagation," in Proceedings of the AAAI conference on artificial intelligence , 2020, vol. 34, no. 04, pp. 5085-5092

2020
[35]

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," in International conference on medical image computing and computer-assisted intervention, 2016: Springer, pp. 424-432

2016
[36]

The multimodal brain tumor image segmentation benchmark (BRATS),

B. H. Menze et al., "The multimodal brain tumor image segmentation benchmark (BRATS)," IEEE transactions on medical imaging, vol. 34, no. 10, pp. 1993-2024, 2014

1993