Post-Operative Glioma Segmentation via Loss Stabilization, Normalization and Subspace Attention

Alexandru Cri\c{s}an; Diana Borza

REVIEW 5 major objections 4 minor 20 references

The paper claims that the standard Generalized Dice Loss destabilizes post-operative glioma segmentation under cross-site shift, while hybrid Dice-cross-entropy with subspace attention stays stable.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 08:21 UTC pith:VVZEREDD

load-bearing objection Plausible and practically relevant GDL-collapse observation, but the external-validation evidence needs label-harmonization proof and significance tests before the headline number is usable. the 5 major comments →

arxiv 2607.22749 v1 pith:VVZEREDD submitted 2026-07-23 cs.CV cs.AI

Post-Operative Glioma Segmentation via Loss Stabilization, Normalization and Subspace Attention

Alexandru Cri\c{s}an , Diana Borza This is my paper

classification cs.CV cs.AI

keywords post-operative glioma segmentationdomain shiftGeneralized Dice Losshybrid Dice-CE losssubspace attentioncontrastive learningintensity normalizationmodel ensembling

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the standard Generalized Dice Loss (GDL), a popular objective for imbalanced medical segmentation, is a poor default for post-operative glioma segmentation when a model trained at one institution is applied at another. On the internal validation set the authors measure 0.88 Whole Lesion Dice, but the same model drops to 0.73 on a held-out external cohort, while a manually weighted hybrid Dice-plus-cross-entropy loss degrades far less. The authors attribute this collapse to GDL's automatic volume-based class weighting, which bakes in the training site's tissue proportions and becomes miscalibrated under domain shift. They then show that two lightweight additions—brain-masked percentile normalization paired with voxel-level contrastive learning, and a Subspace-Aware Class Attention module that recalibrates bottleneck features per tumor class—improve stability and active-tumor sensitivity. Ensembling the stable variants with a separately trained convolutional U-Net brings every stable configuration to 0.94 Whole Lesion Dice, with the best boundary error at 2.92 mm.

Core claim

The paper argues that GDL collapses under domain shift: a model trained with it falls from 0.88 to 0.73 Whole Lesion Dice and its boundary error grows from 11.48 mm to 36.31 mm on a held-out external set. The mechanism is GDL's volume-based class weighting, which encodes the training site's tissue proportions; once those proportions shift, the loss guides the optimizer with a corrupted signal. A manually weighted hybrid Dice + cross-entropy objective removes this prior and trains stably across sites. The paper also introduces Subspace-Aware Class Attention, which splits the bottleneck into four class-aligned subspaces and applies class-conditioned queries per spatial token, raising enhancing

What carries the argument

The central mechanism is the contrast between two loss objectives. Generalized Dice Loss (GDL) weights each class by its training-set volume, so under domain shift the weights encode the wrong prior; the paper's baseline uses a manually weighted hybrid Dice + cross-entropy objective, called DiceWCE, that decouples from volume statistics entirely. Two auxiliary components do the remaining work. The Subspace-Aware Class Attention (SACA) module partitions the bottleneck into S=4 non-overlapping subspaces whose width matches the four segmentation classes, applies shared class-conditioned queries to each subspace per spatial token, and reintegrates the class-calibrated features through a zero-ini

Load-bearing premise

The load-bearing premise is that the held-out external cohort's ground-truth labels and the published comparison numbers use the same four tumor classes and comparable protocols; if those label definitions, scanner mix, or baseline training procedures differ, the measured 0.88-to-0.73 gap is not a clean loss-function effect.

What would settle it

Retrain the GDL variant on the internal cohort after resampling class proportions to match the external cohort's tissue volumes, keeping all images unchanged. If the Whole Lesion Dice gap shrinks toward the hybrid-loss gap, the volume-weighting mechanism is confirmed; if the gap persists, intensity differences or label incompatibility—not GDL—are driving the collapse.

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

If this is right

If GDL genuinely collapses under domain shift, multi-site post-operative segmentation pipelines that default to GDL should be re-evaluated; switching to a static weighted hybrid Dice-CE removes the collapse.
The SACA module increases active-tumor recall by 8% and tumor-core recall by 7% on internal validation, which matters for catching residual or recurrent disease rather than merely optimizing overlap.
Brain-masked percentile normalization plus voxel-level contrastive learning gives stable external Dice and makes ensemble behavior more consistent, reaching 0.94 Whole Lesion Dice with the lowest variance.
Ensembling helps only when the component model is stable: a GDL-trained variant pulls the combined model down from 0.94 to 0.90 Whole Lesion Dice and from 3.03 mm to 7.70 mm boundary error, showing that a biased model is an actively harmful ensemble member.
Cross-site generalization gaps in this setting are large enough that internal validation alone substantially overstates expected clinical performance.

Where Pith is reading between the lines

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

The paper does not test whether the GDL collapse is caused specifically by volume weighting or by intensity mismatch; one could retrain the GDL variant on a synthetic version of the internal dataset whose class proportions are rebalanced to match the external cohort, and see whether the gap closes.
The SACA module sets the number of subspaces equal to the number of classes (S=K=4); a natural extension is to test how much of the sensitivity gain survives when the subspace count diverges from the class count, especially on datasets with a different taxonomy.
The normalization variant trades a small loss in fine-region overlap (TC and ET Dice) for better boundary and ensemble stability; tuning the contrastive ramp-up schedule might recover that overlap, which the paper leaves as future work.
Because training uses a single site, the safest immediate reading is as a diagnostic checklist—loss stability should be probed before clinical deployment—rather than as a validated multi-site model; the authors state this single-site limitation explicitly.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Plausible and practically relevant GDL-collapse observation, but the external-validation evidence needs label-harmonization proof and significance tests before the headline number is usable.

read the letter

Short version: if you work in medical segmentation, this paper is worth knowing about because it claims — with a clean internal ablation — that GDL, the default loss in many pipelines, collapses on an external post-op glioma cohort while hybrid Dice-CE does not. That is a useful empirical data point. The rest is incremental: SACA is a new attention module but combines known ideas; percentile normalization and voxel contrastive learning are standard enough.

What the paper does well: the ablation is properly controlled along three axes, hyperparameters are specified, they report medians/IQRs and per-region metrics, and the discussion acknowledges the trade-offs (SACA and Norm slightly hurt TC/WL overlap while improving sensitivity). The negative ensemble result — GDL pulling nnU-Net down — is genuinely interesting and consistent with their explanation. The main limitation statement is honest: single-site training only.

Where it gets soft. The load-bearing generalization claim is not yet load-bearing. Section 3.1 describes the UCSF-ALPTDG subset only as a 100-scan/50-patient set with no label-harmonization step, yet Table 2 reports per-class Dice for all four classes. We don’t know that UCSF ground truth uses the same ET/NETC/RC/SNFH taxonomy and annotation protocol. If labels differ, the 0.88-to-0.73 drop could be partly label mismatch, not domain shift. The ‘official nnU-Net results from [4]’ were adopted without replicating that model on the same split/preprocessing, so the in-distribution reference is wobbly. There are no significance tests anywhere; the internal GDL vs Baseline WL difference (0.88 vs 0.91) has overlapping SDs, and the ET sensitivity gain is a bar chart without error bars. The ensemble headline (2.92 vs 3.03 mm HD95) is within noise. No code or data release, which makes the central claim harder to vet.

I think the stress-test concern lands: the magnitude of the generalization gap is not as clean as claimed. That said, the direction is plausible and the internal evidence is mostly consistent. With a label-comparability analysis, significance tests or error bars, and a replicated or clearly defined nnU-Net baseline, this could be a solid paper. For now it is a strong candidate with shaky external scaffolding.

Bottom line: send to peer review — the question is important enough to spend referee time on — but the referees should push hard on the UCSF labels and the borrowed baseline. I would bring it to a reading group for the methodology discussion, though I would not cite the headline numbers yet.

Referee Report

5 major / 4 minor

Summary. The paper studies post-operative glioma segmentation with SwinUNETR on MU-GLIOMA-POST (training/internal validation) and UCSF-ALPTDG (external validation). It compares a Generalized Dice Loss (GDL) configuration against a hybrid Dice+CE baseline, and proposes two refinements: a brain-masked percentile-normalization plus voxel-level contrastive-loss variant, and a Subspace-Aware Class Attention (SACA) module at the bottleneck. The central empirical claim is that GDL collapses under domain shift (WL Dice 0.88→0.73, HD95 11.48→36.31 mm), while stable variants, when ensembled with nnU-Net, reach WL Dice 0.94. The authors also report that SACA improves enhancing-tumor sensitivity by 8% and that the SACA ensemble achieves the best boundary error (HD95 2.92 mm).

Significance. If the GDL-collapse result is valid, the paper makes a practically important point: the default region-based loss in many post-operative segmentation pipelines can be a poor choice for multi-site deployment. The external UCSF evaluation is a genuine out-of-distribution test, and the ablation structure—loss, normalization, attention—is sensible. The proposed hybrid Dice-CE loss and SACA module are lightweight and plausibly useful, and the paper is honest about its single-site training limitation. However, the current evidence is not yet strong enough: the external label comparability is unverified, no significance tests are reported, the SACA sensitivity gain is shown without error bars, and the normalization/contrastive variant does not improve aggregate external metrics over baseline. The central message is interesting but needs substantial additional validation.

major comments (5)

[§3.1, Table 2] The load-bearing claim that GDL collapses under domain shift (WL Dice 0.88→0.73) presupposes that the UCSF-ALPTDG 100-scan subset has the same four-class label semantics as MU-GLIOMA-POST. Section 3.1 describes the subset only as a '100-scan (50 patients)' set and gives no label-harmonization step, no annotation-protocol statement, and no verification that each scan contains all four classes (ET, NETC, SNFH, RC). If UCSF annotations use a different definition of RC or SNFH, or omit classes, the reported gap could be dominated by label mismatch. In addition, the 'official nnU-Net results from [4]' adopted in Table 2 as an in-distribution reference are not shown to come from the same 100-scan split, the same preprocessing, or the same four-class output. Please provide explicit label-compatibility evidence (e.g., label statistics, protocol description, qualitative agreement) and either repr
[§4.2, Table 2] No significance testing is reported for any cross-configuration comparison. The central GDL-collapse result is supported by WL Dice 0.73±0.21 vs Baseline 0.84±0.17 (n=100); given the large SDs, the difference is not self-evidently reliable. The same issue affects the internal comparisons in Table 1 (e.g., GDL Variant WL 0.88±0.16 vs Baseline 0.91±0.19). Report paired or unpaired statistical tests with appropriate multiple-comparison correction, or at least confidence intervals for the differences in WL, TC, ET, NETC, SNFH, RC, and HD95. Without this, the claim that GDL is 'unstable' rather than merely slightly worse is not quantified.
[§4.3, Fig. 2] The SACA module's main claimed benefit—8% absolute / 9.1% relative improvement in ET sensitivity—is presented as a bar chart with no error bars, no subject-level variance, and no significance test. Table 1 shows that SACA's ET Dice (0.83±0.21) is not above Baseline (0.84±0.21), so the sensitivity gain could be offset by precision loss. Please report the full precision-recall trade-off (e.g., sensitivity and precision, or precision-recall curves) with per-subject variability and statistical significance. Without this, the SACA contribution is not established.
[§3.4, §4.2] The GDL Variant is described only as 'standard GDL' (no equation, exact class-weight scheme, or hyperparameters are given), and it is compared against a manually weighted hybrid DiceWCE with λ=0.6. It is therefore unclear whether the observed collapse is intrinsic to the GDL loss family or to a particular untuned implementation. Please specify the GDL form precisely and, ideally, include a second GDL configuration (e.g., with different weight normalization or smoothing) to demonstrate that the instability is not an artifact of one hyperparameter setting. The fair-comparison concern is exacerbated by the absence of hyperparameter sensitivity analysis for λ, τ, and S.
[§3.5, Table 2] The normalization+contrastive variant is claimed to improve cross-site consistency, but in Table 2 the Norm Variant's external WL Dice (0.83±0.18) is numerically below Baseline (0.84±0.17) and its WL HD95 (13.61±17.24) is worse than Baseline (10.04±16.48); only NETC improves. Thus the stated benefit is not visible on the aggregate WL/TC/ET metrics. In addition, the contribution is confounded: Norm Variant changes both the input normalization and the training objective, so the effect cannot be attributed to either component. Please report the two changes separately and clarify whether the claim is limited to NETC. Also give the actual values of e_warm and e_ramp and state whether λ=0.6, τ=0.5, and S=4 were tuned or chosen a priori.

minor comments (4)

[§3.3, Fig. 1] Notation is inconsistent: the text uses f∈R^{B×N×C}, while Fig. 1 uses 'B×Nb×768' and 'f1 B×Nb×192'; the 'LN' layer is never defined; the softmax subscript in the attention equation is ambiguous. Please clean up the equation formatting and define all symbols.
[§3.5] The phrase 'hardware-invariant' presumably means 'scanner-invariant' or 'acquisition-protocol-invariant'; the warmup/ramp values e_warm and e_ramp are used in the formula for α_contra but their numerical values are never reported.
[§4.4, Table 3] The statement that 'every stable configuration' reaches WL Dice 0.94 is overstated because standalone nnU-Net already achieves 0.94; the ensemble adds little on Dice. The HD95 difference between SACA ensemble (2.92 mm) and nnU-Net (3.03 mm) is far smaller than the reported SDs (5.17 vs 5.49 mm), so the claim of 'best boundary fidelity' should be qualified.
[General] Minor typos and naming inconsistencies: 'Cris,an' in the author line, 'UCSF-ALPTDG' vs 'UCSF' used interchangeably, Figure 3's color legend for labels is only in the caption but the caption does not mention RC color; reference [4] is cited both as the dataset source and as the source of the nnU-Net scores, which should be clarified.

Circularity Check

0 steps flagged

No circularity found; the paper's central claims are empirical ablations on held-out data.

full rationale

The paper's load-bearing claims are all experimental measurements, not quantities derived from their own definitions or from self-citations. The GDL-collapse claim (WL Dice dropping from 0.88 internal to 0.73 external) is an observed result of training a GDL Variant and evaluating it on a held-out external dataset; no fitted parameter is renamed as a prediction. The SACA module's ET-sensitivity improvement is a learned effect reported from validation-set metrics, not an identity guaranteed by the module's construction. The normalization and contrastive variants are likewise ablated empirically, with hyperparameters (loss weight 0.6, contrastive temperature 0.5, warmup/ramp schedules) stated as inputs rather than inferred from the target metrics. The paper cites prior work for losses, normalization, contrastive learning, SwinUNETR, and nnU-Net, but does not rely on a self-citation chain to justify its central conclusion. The correctness concern noted by the reader's take — that UCSF label comparability and the borrowed nnU-Net baseline from [4] may be unverified — is a question of experimental validity and external reproducibility, not circularity. No equation in the paper reduces to its own input, and no prediction is forced by construction. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claims rest on standard ML hyperparameters, not physical parameters. No fitted constants are masked as predictions; the main free choices are the loss balance (0.6), contrastive temperature/voxel budgets/ramp schedule, class weights, and S=K=4. The largest unverified external assumption is UCSF label/taxonomy compatibility and the borrowed nnU-Net baseline.

free parameters (6)

DiceWCE CE weight λ=0.6 = 0.6
Chosen by hand; controls the hybrid loss balance and is not swept in the ablation.
Class-weight vector in DiceWCE = not specified
Active-tumor subclasses get higher weights but actual values are not reported; affects loss and results.
Contrastive temperature τ = 0.5
Chosen manually; controls contrastive pair scaling.
Positive/negative voxel counts = 64/128
Sampling budget for contrastive pairs; hand-set.
Contrastive ramp epochs e_warm/e_ramp = not specified
Defined but never given values; determines when the contrastive loss activates.
Subspace count S=4 = 4
Set equal to number of classes; no sensitivity analysis.

axioms (5)

domain assumption The BraTS 2024 four-class taxonomy (ET/NETC/SNFH/RC) applies to both MU-GLIOMA-POST and UCSF-ALPTDG ground truth.
Section 2 adopts taxonomy; Sec 3.1/4.2 assumes UCSF labels use it without verification.
domain assumption The official nnU-Net scores from [4] are a valid in-distribution reference for the external UCSF-ALPTDG set.
Table 2 uses adopted baselines; training protocol may differ from the paper's own external evaluation.
domain assumption Brain-masked percentile clipping (1st–99th) plus Z-score aligns intensity distributions across sites.
Sec 3.1; no empirical proof beyond results.
domain assumption Voxel-level contrastive loss promotes hardware invariance of bottleneck features.
Sec 3.5; the mechanism is assumed, not measured.
domain assumption SwinUNETR bottleneck tokens can be partitioned into S=4 subspaces that correspond to classes.
Sec 3.3; structural assumption of SACA.

invented entities (1)

Subspace-Aware Class Attention (SACA) module no independent evidence
purpose: Re-calibrates bottleneck feature channels per class via shared class-conditioned queries
No external validation/falsifiable handle; supported only by internal ablation in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 9107 in / 10845 out tokens · 90510 ms · 2026-08-01T08:21:53.359328+00:00 · methodology

0 comments

read the original abstract

Tracking residual tumor after surgery is essential for catching recurrence early, but automating post-operative glioma segmentation remains a difficult task. Although transformer-based architectures, such as SwinUNETR, achieved impressive results, few studies test how well they generalize across clinical protocols. In this paper, we conduct an ablation study on the MU-GLIOMA-POST and UCSF-ALPTDG datasets and show that the standard Generalized Dice Loss (GDL) is unstable under domain shift: the Whole Lesion (WL) Dice drops from 0.88 on the internal validation set to 0.73 on the external UCSF test set. To address this, we pair brain-masked percentile normalization with voxel-level contrastive learning. We also propose a Subspace-Aware Class Attention (SACA) module that re-calibrates the bottleneck features and raises Enhancing Tumor (ET) sensitivity by 8% (9.1% relative improvement) on internal validation. Ensembling these refinements with nnU-Net brings every stable configuration to a WL Dice of 0.94, and the SACA variant ensemble achieves the best boundary error (HD95) of 2.92 mm on MU-GLIOMA-POST.

Figures

Figures reproduced from arXiv: 2607.22749 by Alexandru Cri\c{s}an, Diana Borza.

**Figure 1.** Figure 1: Overview of the SACA module. The bottleneck features are partitioned into S=4 latent subspaces for class-conditioned attention using shared projections and global queries Qcls. The final output f ′ is integrated through a learned gated residual connection: f ′ = f + σ(γ) · O. 3.4 Loss Function Ablation: GDL vs. DiceWCE A main hypothesis of this study is that the standard generalized dice loss (GDL) [17] b… view at source ↗

**Figure 2.** Figure 2: Sensitivity (recall) analysis across tumor-active regions. The bar chart illustrates the systematic improvement in detection performance provided by the SACA Variant compared to the Baseline SwinUNETR [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative results. Columns (left to right): T1c axial slice, Ground Truth (GT), Baseline, Norm Variant, and SACA Variant. Segmentation labels: ET (blue), TC (yellow), RC (green), and SNFH (red). Case 1: WL Dice > 0.95; Case 2: WL Dice ≈ 0.80; Case 3: morphology featuring a large resection cavity. corrupt the stronger model’s posterior estimates, making a biased model an actively harmful ensemble componen… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 3 linked inside Pith

[1]

Neuro-Oncology Ad- vances6(1), vdae199 (2024)

Cepeda, S., Romero, R., Luque, L., García-Pérez, D., Blasco, G., Luppino, L.T., Kuttner, S., Esteban-Sinovas, O., Arrese, I., Solheim, O., et al.: Deep learning- based postoperative glioblastoma segmentation and extent of resection evaluation: Development, external validation, and model comparison. Neuro-Oncology Ad- vances6(1), vdae199 (2024)

2024
[2]

arXiv preprint arXiv:2102.04306 (2021)

Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

Pith/arXiv arXiv 2021
[3]

Neurotherapeutics14(2), 307– 320 (2017)

Ellingson, B.M., Wen, P.Y., Cloughesy, T.F.: Modified criteria for radiographic response assessment in glioblastoma clinical trials. Neurotherapeutics14(2), 307– 320 (2017)

2017
[4]

Radiology: Artificial Intelligence6(4), e230182 (2024)

Fields, B.K., Calabrese, E., Mongan, J., Cha, S., Hess, C.P., Sugrue, L.P., Chang, S.M., Luks, T.L., Villanueva-Meyer, J.E., Rauschecker, A.M., et al.: The university of california san francisco adult longitudinal post-treatment diffuse glioma mri dataset. Radiology: Artificial Intelligence6(4), e230182 (2024)

2024
[5]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

2021
[6]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

He, Y., Nath, V., Yang, D., Tang, Y., Myronenko, A., Xu, D.: Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 416–426. Springer (2023) 12 A. Cris ,an and D. Borza

2023
[7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Holden Helland, R., Bouget, D., Eijgelaar, R.S., De Witt Hamer, P.C., Barkhof, F., Solheim, O., Reinertsen, I.: Glioblastoma segmentation from early post-operative mri: Challenges and clinical impact. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 284–294. Springer (2024)

2024
[8]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

2018
[9]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021
[10]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 488–498. Springer (2024)

2024
[11]

Advances in neural information pro- cessing systems30(2017)

Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information pro- cessing systems30(2017)

2017
[12]

Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

2021
[13]

Scientific data12(1), 1847 (2025)

Mahmoud, E., Gass, J., Dhemesh, Y., Greaser, J., Pogorzelski, K., Isufi, E., Gar- rett, F., Thacker, J., Tahon, N.h., Sinclair, J., et al.: Mu-glioma post: A comprehen- sive dataset of automated mr multi-sequence segmentation and clinical features. Scientific data12(1), 1847 (2025)

2025
[14]

arXiv preprint arXiv:1804.03999 (2018)

Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

Pith/arXiv arXiv 2018
[15]

In: Proceedings of SPIE–the Inter- national Society for Optical Engineering

Reinhold, J.C., Dewey, B.E., Carass, A., Prince, J.L.: Evaluating the impact of intensity normalization on mr image synthesis. In: Proceedings of SPIE–the Inter- national Society for Optical Engineering. vol. 10949, p. 109493H (2019)

2019
[16]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015
[17]

In: International Workshop on Deep Learning in Medical Image Analysis

Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: International Workshop on Deep Learning in Medical Image Analysis. pp. 240–
[18]

arXiv preprint arXiv:2405.18368 (2024)

de Verdier, M.C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N.H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., et al.: The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. arXiv preprint arXiv:2405.18368 (2024)

Pith/arXiv arXiv 2024
[19]

Computerized Medical Imaging and Graphics95, 102026 (2022)

Yeung, M., Sala, E., Schönlieb, C.B., Rundo, L.: Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image seg- mentation. Computerized Medical Imaging and Graphics95, 102026 (2022)

2022
[20]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

You,C.,Zhao,R.,Staib,L.H.,Duncan,J.S.:Momentumcontrastivevoxel-wiserep- resentation learning for semi-supervised volumetric medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 639–652. Springer (2022)

2022

[1] [1]

Neuro-Oncology Ad- vances6(1), vdae199 (2024)

Cepeda, S., Romero, R., Luque, L., García-Pérez, D., Blasco, G., Luppino, L.T., Kuttner, S., Esteban-Sinovas, O., Arrese, I., Solheim, O., et al.: Deep learning- based postoperative glioblastoma segmentation and extent of resection evaluation: Development, external validation, and model comparison. Neuro-Oncology Ad- vances6(1), vdae199 (2024)

2024

[2] [2]

arXiv preprint arXiv:2102.04306 (2021)

Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

Pith/arXiv arXiv 2021

[3] [3]

Neurotherapeutics14(2), 307– 320 (2017)

Ellingson, B.M., Wen, P.Y., Cloughesy, T.F.: Modified criteria for radiographic response assessment in glioblastoma clinical trials. Neurotherapeutics14(2), 307– 320 (2017)

2017

[4] [4]

Radiology: Artificial Intelligence6(4), e230182 (2024)

Fields, B.K., Calabrese, E., Mongan, J., Cha, S., Hess, C.P., Sugrue, L.P., Chang, S.M., Luks, T.L., Villanueva-Meyer, J.E., Rauschecker, A.M., et al.: The university of california san francisco adult longitudinal post-treatment diffuse glioma mri dataset. Radiology: Artificial Intelligence6(4), e230182 (2024)

2024

[5] [5]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

2021

[6] [6]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

He, Y., Nath, V., Yang, D., Tang, Y., Myronenko, A., Xu, D.: Swinunetr-v2: Stronger swin transformers with stagewise convolutions for 3d medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 416–426. Springer (2023) 12 A. Cris ,an and D. Borza

2023

[7] [7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Holden Helland, R., Bouget, D., Eijgelaar, R.S., De Witt Hamer, P.C., Barkhof, F., Solheim, O., Reinertsen, I.: Glioblastoma segmentation from early post-operative mri: Challenges and clinical impact. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 284–294. Springer (2024)

2024

[8] [8]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

2018

[9] [9]

Nature methods18(2), 203–211 (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

2021

[10] [10]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Isensee, F., Wald, T., Ulrich, C., Baumgartner, M., Roy, S., Maier-Hein, K., Jaeger, P.F.: nnu-net revisited: A call for rigorous validation in 3d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 488–498. Springer (2024)

2024

[11] [11]

Advances in neural information pro- cessing systems30(2017)

Lakshminarayanan, B., Pritzel, A., Blundell, C.: Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information pro- cessing systems30(2017)

2017

[12] [12]

Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer:Hierarchical visiontransformerusingshiftedwindows.In:Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)

2021

[13] [13]

Scientific data12(1), 1847 (2025)

Mahmoud, E., Gass, J., Dhemesh, Y., Greaser, J., Pogorzelski, K., Isufi, E., Gar- rett, F., Thacker, J., Tahon, N.h., Sinclair, J., et al.: Mu-glioma post: A comprehen- sive dataset of automated mr multi-sequence segmentation and clinical features. Scientific data12(1), 1847 (2025)

2025

[14] [14]

arXiv preprint arXiv:1804.03999 (2018)

Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)

Pith/arXiv arXiv 2018

[15] [15]

In: Proceedings of SPIE–the Inter- national Society for Optical Engineering

Reinhold, J.C., Dewey, B.E., Carass, A., Prince, J.L.: Evaluating the impact of intensity normalization on mr image synthesis. In: Proceedings of SPIE–the Inter- national Society for Optical Engineering. vol. 10949, p. 109493H (2019)

2019

[16] [16]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015

[17] [17]

In: International Workshop on Deep Learning in Medical Image Analysis

Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge Cardoso, M.: Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In: International Workshop on Deep Learning in Medical Image Analysis. pp. 240–

[18] [18]

arXiv preprint arXiv:2405.18368 (2024)

de Verdier, M.C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N.H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., et al.: The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. arXiv preprint arXiv:2405.18368 (2024)

Pith/arXiv arXiv 2024

[19] [19]

Computerized Medical Imaging and Graphics95, 102026 (2022)

Yeung, M., Sala, E., Schönlieb, C.B., Rundo, L.: Unified focal loss: Generalising dice and cross entropy-based losses to handle class imbalanced medical image seg- mentation. Computerized Medical Imaging and Graphics95, 102026 (2022)

2022

[20] [20]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

You,C.,Zhao,R.,Staib,L.H.,Duncan,J.S.:Momentumcontrastivevoxel-wiserep- resentation learning for semi-supervised volumetric medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 639–652. Springer (2022)

2022