Multi-step Cascaded Networks for Brain Tumor Segmentation

Gongning Luo; Kuanquan Wang; Xiangyu Li

REVIEW 3 major objections 6 minor 14 references

Multi-step Cascaded Networks for Brain Tumor Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Cascaded 3D U-nets hit 0.886 whole-tumor Dice on BraTS 2019.

desk verdict A credible BraTS 2019 cascade method with real validation scores, but its central advantage claim rests on no baseline at all. read the letter →

arxiv 1908.05887 v3 pith:NCMYR45U submitted 2019-08-16 eess.IV cs.CV

Xiangyu Li , Gongning Luo , Kuanquan Wang This is my paper · ORCID

classification eess.IVcs.CV

keywords braintumorsegmentationcascadednetwork3DU-Netdeepsupervisionfocallosscoarse-to-fineBraTS2019MRI

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that brain tumor substructures—whole tumor, tumor core, and enhancing tumor—are best segmented in a coarse-to-fine cascade that respects their nested containment. A first 3D U-net segments the whole tumor; its predicted mask is multiplied into the MRI volume as spatial prior for a second step that segments the tumor core, and that mask guides a third step for the enhancing tumor. The three steps are trained end-to-end with deep supervision and focal loss to cope with vanishing gradients and extreme class imbalance. On the BraTS 2019 validation set the method reports mean Dice coefficients of 0.886, 0.813, and 0.771 for the three structures. If correct, this gives a working automatic glioma segmentation pipeline on the standard benchmark.

What carries the argument

The central mechanism is the multi-step cascade with mask multiplication. In each finer step, the previous step's predicted binary mask is multiplied voxel-wise into the chosen MRI modality before it enters the next 3D U-net, so the network concentrates on the remaining candidate region; this acts as a spatial-attention prior consistent with the nested containment hierarchy of whole tumor, tumor core, and enhancing tumor. The auxiliary deep-supervision outputs in the expanding pathway propagate gradients to earlier layers, and focal loss reweights easy versus hard voxels to counter the strong background-foreground imbalance.

What would settle it

Replace the predicted whole-tumor mask with the ground-truth whole-tumor mask during inference on the validation set. If the tumor-core and enhancing-tumor Dice scores improve substantially, the cascade's accuracy is bottlenecked by first-step recall; if they stay the same, error propagation from the first step is not the limiting factor. The paper itself reports no ablation of this kind.

Watch

Extended reading notes

Core claim

The central claim is that converting the multi-class tumor segmentation problem into three linked binary problems, ordered by the anatomical hierarchy, improves accuracy over a single network that must separate all structures at once. Each cascade step is a 3D U-net with three auxiliary deep-supervision outputs; the loss is focal loss. The chosen modalities mirror the manual annotation protocol: Flair and T1ce for whole tumor, T1ce alone for tumor core and enhancing tumor. The authors report mean Dice of 0.886 for whole tumor, 0.813 for tumor core, and 0.771 for enhancing tumor on the BraTS 2019 validation data, and note that the most common failure is predicting the whole tumor as tumor core, especially for LGG cases.

Load-bearing premise

The whole pipeline assumes the first step's whole-tumor mask contains the true tumor core and enhancing tumor; if that first mask misses any tumor voxels, the two finer steps never get a chance to find them.

Editorial extensions

If this is right

The cascade turns one hard multi-class segmentation problem into three simpler binary problems, which should generalize better when labeled training data are limited.
Because each step sees only the voxels inside the previous mask, the final enhancing-tumor prediction is constrained to lie inside the predicted tumor core, enforcing anatomical consistency by construction.
End-to-end training outperformed a curriculum in which the steps were trained one at a time, suggesting joint parameter updates help the cascade fit the training data.
Weighting the three cascade steps up or down changed final results little, so the method is insensitive to that hyperparameter.
The observed confusion of whole tumor for tumor core points to inter-volume intensity variation as a remaining source of error.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An implication the paper leaves untested: the pipeline's ceiling is set by the recall of the first-step whole-tumor mask; any voxel missed there is unrecoverable by the tumor-core and enhancing-tumor steps. A direct test would be to feed ground-truth whole-tumor masks into the later steps and measure how much the tumor-core and enhancing-tumor Dice improve.
The hard mask multiplication could be replaced by a soft probability map from the previous step, which would let gradients flow across steps and might allow the finer networks to correct small whole-tumor misses; the paper does not explore this.
The same coarse-to-fine cascade with per-step modality selection could transfer to other nested lesion structures, such as pancreas regions or prostate zones, wherever a containment hierarchy is known a priori.
Because the authors observe worse variance on tumor core than enhancing tumor and attribute it to LGG cases, a stratified analysis by tumor grade, or a histogram-equalization preprocessing step, is a concrete next experiment suggested by their own discussion.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is a straightforward BraTS 2019 method report: three 3D U-nets in a hard cascade, deep supervision, focal loss, and per-step modality selection. What's genuinely new is the specific combination and the validation results, which come from the official BraTS 2019 evaluation platform and are internally consistent (WT > TC > ET, with sensitivity following the same order). A code link is provided, so the empirical part is reproducible in principle. That is real value, and the paper does not overstate the absolute scores.

The soft spot is the causal claim. Section 2.1 multiplies the previous mask into the next input, and because TC and ET are strict subsets of WT, any WT false negative is locked out of the finer steps. The paper reports no one-step baseline, no oracle-mask ablation, and no analysis of where the WT false negatives fall relative to TC/ET ground truth. So the Conclusion's phrase 'proved to be more effective than the one-step counterpart' is unsupported. This is not a refutation of the method; a cascade can still be useful even if it cannot recover WT misses. But the paper does not demonstrate that benefit, and the hard-masking design makes that omission consequential rather than cosmetic.

Other issues are minor but worth noting: no error bars, no exact hyperparameters for the focal loss (alpha_t and gamma are never given), and the Discussion mentions curriculum learning and step-weighting experiments without any quantitative results. Those are correctable omissions. The citation pattern is fine: Havaei et al. is credited for cascaded brain tumor segmentation, and the BraTS benchmark references are standard. Self-citation is not a concern here.

Who is this for? Someone building a brain tumor segmentation pipeline and looking for a compact, working recipe with plausible numbers. It is not a methods paper that establishes why the cascade works; it is an engineering report. A serious referee would find real value in the empirical results but should push for a proper ablation and baseline comparison before the comparative claims are accepted.

Recommendation: send it out. The work deserves reviewer time, primarily to demand the missing controls, not because the architecture is theoretically deep. My own verdict on the paper as written is 'revise with ablations', not 'reject'.

Referee Report

3 major / 6 minor

Summary. The paper proposes a multi-step cascaded 3D U-Net for glioma substructure segmentation on BraTS 2019. The pipeline first segments the whole tumor (WT) from Flair and T1ce, then segments the tumor core (TC) and enhancing tumor (ET) from T1ce with the previous mask multiplied into the input as prior information; all steps are trained end-to-end with deep supervision and focal loss. On the official BraTS 2019 validation platform the authors report mean Dice coefficients of 0.886 for WT, 0.813 for TC, and 0.771 for ET, with code publicly available. The paper also describes N4 bias-field correction, random cropping, data augmentation, and a curriculum-learning comparison that reportedly showed no benefit over end-to-end training.

Significance. If the reported scores are accurate, the paper describes a competitive automatic glioma segmentation pipeline on a standard benchmark; the use of the official evaluation platform and the release of code are strengths. However, the paper's central causal claims—that the cascade itself is more effective than a one-step network and that per-step modality selection is beneficial—are not backed by any ablation in the manuscript. The descriptive results are plausible, but the attribution of the scores to the cascade and modality choices is an unsupported step that can be remedied with targeted experiments.

major comments (3)

[Section 2.1, Figure 2, and Conclusion] The conclusion states that the multi-step network 'has proved to be more effective than the one-step counterpart,' but the manuscript reports no one-step baseline or three-independent-networks baseline trained with the same architecture and data. Because the TC and ET steps multiply the previous WT mask into the input, any WT false negative (validation WT sensitivity is 0.921 in Table 1) removes those voxels from the reach of the later steps. The manuscript should report an ablation that either (a) feeds oracle ground-truth WT masks into steps 2 and 3 to measure the ceiling imposed by the WT step, or (b) trains a single multi-class network and/or three independent binary networks under identical conditions. Without such an experiment, the TC/ET Dice cannot be attributed to the cascade mechanism rather than to the WT step's recall.
[Section 2.1 and Conclusion] The conclusion also claims that choosing specific modalities for each step 'improves our result a lot compared to the method to use all modalities,' but no experiment comparing modality subsets is reported anywhere in the paper. This is a load-bearing claim for the method's design; an ablation using all four modalities in each cascade step, or in the one-step baseline, is needed before the claim can be accepted.
[Section 3.3 and Table 1] No analysis is provided of the spatial overlap between WT false negatives and true TC/ET voxels. Since TC and ET are strict subsets of WT, even a modest WT sensitivity loss could disproportionately remove enhancing-tumor voxels, and this is exactly the failure mode that the paper needs to rule out. I recommend quantifying the fraction of true TC/ET voxels that fall inside WT false-negative regions on the validation set, and relating that quantity to the observed TC/ET Dice.

minor comments (6)

[Section 3.2] The text 'adding guassion blur' should read 'adding Gaussian blur.'
[Section 3.3] The word 'mothed' should be 'method.'
[Conclusion] The word 'mamual' should be 'manual.'
[Section 2.2] Please define how the three auxiliary outputs are combined with the main output in the total loss, including the weighting coefficients, and provide the values of alpha_t and gamma used in Eq. (1).
[Introduction and References] The phrase 'Inspired by Havaei and iek' should name Cicek et al.; the intended citation to reference [4] is otherwise unclear.
[Figure 3] The architectural details are not fully specified: the number of feature maps, downsampling factors, and the exact locations where auxiliary outputs are attached should be stated to make the network reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claims are external validation scores on a fixed benchmark; the cascade's coarse-to-fine masking is the method under test, not an input-dependent derivation.

full rationale

The paper reports an empirical segmentation method evaluated on the BraTS 2019 benchmark with online platform metrics. There is no derivation chain in which an output is defined in terms of fitted inputs, no parameter fitted to a subset and then renamed as a prediction, and no load-bearing self-citation. The multi-step cascade uses each previous mask as an additional input to the next step; this is the proposed architecture under evaluation, not a circular argument. The absence of an oracle-mask ablation is a legitimate experimental limitation concerning whether the cascade causes the observed performance, but it does not make the reported Dice scores circular. Self-citations in the introduction are background references, not justifications for the paper's conclusions. Therefore no specific circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical engineering result, not a mathematical derivation. It depends on the trustworthiness of the BraTS 2019 labels, the assumed nested structure of tumor substructures, and a set of hand-chosen hyperparameters that are only partially reported. No new theoretical entities are introduced.

free parameters (8)

initial learning rate = 0.001, decayed to 0.0005
Adam optimizer learning rate schedule chosen in Section 3.2; affects convergence and final metrics.
random crop size = 96 x 96 x 96
Sub-volume size in Section 3.2, chosen due to GPU memory; determines context available to the network.
batch size = 1
Set to 1 in Section 3.2; affects gradient noise and normalization behavior.
number of training epochs = 50
Training duration in Section 3.2; stopping is not based on validation performance.
focal loss alpha_t = not reported
Equation (1) defines alpha_t but the value is never given; this changes the loss balance between positive and negative samples.
focal loss gamma = not reported
Equation (1) defines the focusing parameter gamma, but the value used in training is not stated.
data augmentation settings = not reported
Section 3.2 lists rotation, flipping, and Gaussian blur, but probabilities, angle ranges, and blur parameters are omitted.
3D U-net architecture dimensions = not specified in text
The network is shown only as a figure; filter counts, number of downsampling levels, and upsampling details are not stated in the text.

assumptions (4)

domain assumption BraTS 2019 ground truth segmentations are accurate and follow a consistent annotation protocol
Section 3.1 relies on official manual labels as ground truth for training and evaluation.
domain assumption The hierarchical containment of tumor substructures is a valid prior for segmentation
Section 2.1 builds the cascade on the assumption that whole tumor contains tumor core contains enhancing tumor; the paper does not test cases where containment fails.
domain assumption Preprocessed multimodal MRI with N4 correction and z-score normalization is sufficient input for the network
Section 3.1 applies co-registration, skull-stripping, N4 bias correction, and normalization as given, without validating each preprocessing choice.
ad hoc to paper Mask multiplication propagates only useful prior information into the next cascade step
Section 2.1 masks T1ce with the previous whole-tumor output; this design choice is asserted, not derived or ablated, and it can propagate false negatives.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-step Cascaded Networks for Brain Tumor Segmentation." pith.science (2026). https://pith.science/paper/NCMYR45U

@misc{pith2026190805887,
  author       = {Pith},
  title        = {Pith review of: Multi-step Cascaded Networks for Brain Tumor Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NCMYR45U}},
  note         = {Machine review of arXiv:1908.05887}
}

read the original abstract

Automatic brain tumor segmentation method plays an extremely important role in the whole process of brain tumor diagnosis and treatment. In this paper, we propose a multi-step cascaded network which takes the hierarchical topology of the brain tumor substructures into consideration and segments the substructures from coarse to fine .During segmentation, the result of the former step is utilized as the prior information for the next step to guide the finer segmentation process. The whole network is trained in an end-to-end fashion. Besides, to alleviate the gradient vanishing issue and reduce overfitting, we added several auxiliary outputs as a kind of deep supervision for each step and introduced several data augmentation strategies, respectively, which proved to be quite efficient for brain tumor segmentation. Lastly, focal loss is utilized to solve the problem of remarkably imbalance of the tumor regions and background. Our model is tested on the BraTS 2019 validation dataset, the preliminary results of mean dice coefficients are 0.886, 0.813, 0.771 for the whole tumor, tumor core and enhancing tumor respectively. Code is available at https://github.com/JohnleeHIT/Brats2019

Figures

Figures reproduced from arXiv: 1908.05887 by the authors.

**Figure 2.** Overview of the proposed multi-step cascaded network [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Schematic of the 3D U-net architecture with deep supervisions [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Boxplot of the overall performance on both training and validation [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Segmentation result of the whole tumor (WT), Tumor core (TC) and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Segmentation result of the whole tumor (WT), Tumor core (TC) and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 11 canonical work pages

[1]

Proceedings MICCAI-BRATS (2014) 36–39

Zikic, D., Ioannou, Y., Brown, M., Criminisi, A.: Segmentation of brain tumor tissues with convolutional neural networks. Proceedings MICCAI-BRATS (2014) 36–39

work page 2014
[2]

In: 2015 International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), IEEE (2015) 306–309

Zhao, L., Jia, K.: Deep feature learning with discrimination mechanism for brain tumor segmentation and diagnosis. In: 2015 International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), IEEE (2015) 306–309

work page 2015
[3]

Medical image analysis 35 (2017) 18–31

Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., Pal, C., Jodoin, P.M., Larochelle, H.: Brain tumor segmentation with deep neural networks. Medical image analysis 35 (2017) 18–31

2017
[4]

In: Interna- tional conference on medical image computing and computer-assisted intervention, Springer (2016) 424–432

C ¸ i¸ cek,¨O., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u- net: learning dense volumetric segmentation from sparse annotation. In: Interna- tional conference on medical image computing and computer-assisted intervention, Springer (2016) 424–432

work page 2016
[5]

Medical image analysis 36 (2017) 61–78

Kamnitsas, K., Ledig, C., Newcombe, V.F., Simpson, J.P., Kane, A.D., Menon, D.K., Rueckert, D., Glocker, B.: Eﬃcient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical image analysis 36 (2017) 61–78

work page 2017
[6]

Pattern Recognition 88 (2019) 90–100

Chen, S., Ding, C., Liu, M.: Dual-force convolutional neural networks for accurate brain tumor segmentation. Pattern Recognition 88 (2019) 90–100

work page 2019
[7]

IEEE transactions on medical imaging 37(8) (2018) 1943–1954

Ma, C., Luo, G., Wang, K.: Concatenated and connected random forests with mul- tiscale patch driven active contour model for automated brain tumor segmentation of mr images. IEEE transactions on medical imaging 37(8) (2018) 1943–1954

work page 2018
[8]

IEEE transactions on medical imaging 34(10) (2014) 1993–2024

Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10) (2014) 1993–2024

work page 2014

Show all 14 references

[9]

In: Proceedings of the IEEE international conference on computer vision

Lin, T.Y., Goyal, P., Girshick, R., He, K., Doll´ ar, P.: Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision. (2017) 2980–2988

2017
[10]

Scientiﬁc data 4 (2017) 170117 11

Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J.S., Freymann, J.B., Farahani, K., Davatzikos, C.: Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features. Scientiﬁc data 4 (2017) 170117 11

2017
[11]

arXiv preprint arXiv:1811.02629 (2018)

Bakas, S., Reyes, M., Jakab, A., Bauer, S., Rempﬂer, M., Crimi, A., Shinohara, R.T., Berger, C., Ha, S.M., Rozycki, M., et al.: Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats c...

2018 arXiv
[12]

The Cancer Imaging Archive 286 (2017)

Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-lgg collection. The Cancer Imaging Archive 286 (2017)

2017
[13]

the cancer imaging archive (2017)

Bakas, S., Akbari, H., Sotiras, A., Bilello, M., Rozycki, M., Kirby, J., Freymann, J., Farahani, K., Davatzikos, C.: Segmentation labels and radiomic features for the pre-operative scans of the tcga-gbm collection. the cancer imaging archive (2017). (2017)

2017
[14]

IEEE transactions on medical imaging 29(6) (2010) 1310

Tustison, N.J., Avants, B.B., Cook, P.A., Zheng, Y., Egan, A., Yushkevich, P.A., Gee, J.C.: N4itk: improved n3 bias correction. IEEE transactions on medical imaging 29(6) (2010) 1310

2010

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Proceedings MICCAI-BRATS (2014) 36–39

Zikic, D., Ioannou, Y., Brown, M., Criminisi, A.: Segmentation of brain tumor tissues with convolutional neural networks. Proceedings MICCAI-BRATS (2014) 36–39

work page 2014

[2] [2]

In: 2015 International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), IEEE (2015) 306–309

Zhao, L., Jia, K.: Deep feature learning with discrimination mechanism for brain tumor segmentation and diagnosis. In: 2015 International Conference on Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), IEEE (2015) 306–309

work page 2015

[3] [3]

Medical image analysis 35 (2017) 18–31

Havaei, M., Davy, A., Warde-Farley, D., Biard, A., Courville, A., Bengio, Y., Pal, C., Jodoin, P.M., Larochelle, H.: Brain tumor segmentation with deep neural networks. Medical image analysis 35 (2017) 18–31

2017

[4] [4]

In: Interna- tional conference on medical image computing and computer-assisted intervention, Springer (2016) 424–432

C ¸ i¸ cek,¨O., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3d u- net: learning dense volumetric segmentation from sparse annotation. In: Interna- tional conference on medical image computing and computer-assisted intervention, Springer (2016) 424–432

work page 2016

[5] [5]

Medical image analysis 36 (2017) 61–78

Kamnitsas, K., Ledig, C., Newcombe, V.F., Simpson, J.P., Kane, A.D., Menon, D.K., Rueckert, D., Glocker, B.: Eﬃcient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical image analysis 36 (2017) 61–78

work page 2017

[6] [6]

Pattern Recognition 88 (2019) 90–100

Chen, S., Ding, C., Liu, M.: Dual-force convolutional neural networks for accurate brain tumor segmentation. Pattern Recognition 88 (2019) 90–100

work page 2019

[7] [7]

IEEE transactions on medical imaging 37(8) (2018) 1943–1954

Ma, C., Luo, G., Wang, K.: Concatenated and connected random forests with mul- tiscale patch driven active contour model for automated brain tumor segmentation of mr images. IEEE transactions on medical imaging 37(8) (2018) 1943–1954

work page 2018

[8] [8]

IEEE transactions on medical imaging 34(10) (2014) 1993–2024

Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Burren, Y., Porz, N., Slotboom, J., Wiest, R., et al.: The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging 34(10) (2014) 1993–2024

work page 2014