Organizations: Faculty of Information Technology, Van Lang School of Technology, Van Lang University, Ho Chi Minh City 70000, Vietnam · Institute of Information Technology and Electrical-Electronics Engineering, Ho Chi Minh City University of Transport, Ho Chi Minh City 70000, Vietnam
Accurate delineation of whole tumor (WT), tumor core (TC), and enhancing tumor (ET) from multimodal magnetic resonance imaging remains challenging under limited annotation, cross-cohort variation, and severe target sparsity. We propose FSS-UBrain, a region-wise one-shot segmentation framework that uses a labeled positive support slice to condition binary query segmentation separately for WT, TC, and ET. Support-derived foreground and background descriptors guide query-feature adaptation, bottleneck interaction, decoder-side reconstruction, and boundary refinement. Episodic training additionally incorporates hard-negative and fully negative queries with empty-query regularization to suppress spurious foreground activation when the selected region is absent. Although inference operates on two-dimensional support--query slice pairs, checkpoint selection, threshold calibration, and final evaluation are performed after volumetric reconstruction. FSS-UBrain is evaluated on a held-out BraTS 2020 split and under target-supported cross-cohort protocols on BraTS 2023 and BraTS-Africa. Cases used as target support are excluded from the query cohorts, and no target-domain fine-tuning or test-time parameter updates are performed. On BraTS 2020, FSS-UBrain achieves volumetric Dice scores of 89.82%, 82.14%, and 77.42% for WT, TC, and ET, respectively, with corresponding 95th-percentile Hausdorff distance (HD95) values of 11.12, 9.01, and 4.46 mm. It also achieves the highest mean Dice and lowest finite-pair mean HD95 point estimates on BraTS 2023 and BraTS-Africa among the compared few-shot methods. These findings support target-conditioned few-shot segmentation while highlighting sensitivity to support selection and cohort-specific variation.
Figures & tables
Figure 1: Representative BraTS slices illustrating ET sparsity and severe slice-level target-presence imbalance.
Figure 2: Overview of the proposed FSS-UBrain model architecture.
Figure 3: Support-conditioned descriptor adaptation at encoder level ℓ . Masked support features form FG and BG descriptors that are query-refined, matched to query features, and used for descriptor gating and FiLM modulation.
Preprocessing or Augmentation Step and Procedure
Purpose
Applied During
Modality organization. Each axial input is formed by stacking FLAIR, T1, T1ce, and T2 as a four-channel slice.
Preserves the standard multimodal BraTS input representation.
Training, validation, and testing
Region-mask construction. Binary WT, TC, and ET masks are generated using the unified region definitions described in Section 3.1 .
Ensures a consistent clinical-region interpretation across datasets.
Training, validation, and testing
Per-case normalization. For each modality, the mean and standard deviation are computed from nonzero voxels across the complete case volume. Only nonzero voxels are standardized, while zero-valued BG remains zero.
Reduces inter-case intensity variation without altering the zero-valued BG convention.
Training, validation, and testing
Spatial resizing. Images are resized to 256×256 using bilinear interpolation; binary masks are resized using nearest-neighbor interpolation.
Provides a fixed input size while avoiding interpolation-induced label mixing.
Training, validation, and testing
Horizontal and vertical flips. Each image–mask pair is independently flipped horizontally and vertically with probability 0.5 .
Increases invariance to in-plane orientation variation.
Training only
Right-angle rotation. With probability 0.5 , an image–mask pair is rotated by 0∘ , 90∘ , 180∘ , or 270∘ .
Improves robustness to in-plane orientation variation while preserving voxel-grid alignment.
Training only
Table 1: Preprocessing and training-time augmentation applied to support and query inputs.
Evaluation Metric and Formula
Main Advantages
Main Limitations
Dice coefficient. This overlap measure is defined as ( Milletari et al., 2016 ; Taha and Hanbury, 2015 ) Dice(Mr,Mr)=2TP+FP+FN+ϵ32TP+ϵ3, (10) where ϵ3=10−6 is the small constant used for numerical stability. Dice is the voxel-wise F1 score of the predicted FG region and is reported as a percentage.
Primary overlap measure for imbalanced binary segmentation; weights FP and FN equally while excluding TN , so it is not inflated by the large BG volume; comparable in scale with the BraTS literature.
Invariant to error location; merges over- and under-segmentation into one value; unstable for sparse regions such as ET; degenerate for jointly empty pairs.
HD95. The 95th-percentile Hausdorff distance is defined as ( Taha and Hanbury, 2015 ) HD95(Mr,Mr)=Q0.95(D(∂Mr,∂Mr)⊎D(∂Mr,∂Mr)), (11) where ⊎ denotes concatenation with multiplicities preserved, Q0.95 returns the 95th percentile of the resulting distance collection, whereas ∂Mr and ∂Mr denote the surfaces of the reference and predicted masks, respectively. For two surfaces A and B , D(A,B) denotes the collection of directed nearest-surface distances from A to B .
Robust surface-distance measure in physical space that reduces sensitivity to isolated outlier voxels; interpretable as a clinical margin; sensitive to the geometric placement of errors that overlap metrics cannot express.
Undefined when exactly one mask is empty, yielding a finite-pair summary over a varying subset; conveys no volumetric information; depends on voxel spacing; the discarded 5% tail may hide extreme deviations.
Intersection over union (IoU). The overlap between prediction and reference is ( Taha and Hanbury, 2015 ) IoU(Mr,Mr)=TP+FP+FNTP. (12) Also known as the Jaccard index, it satisfies IoU=Dice/(2−Dice) for an individual volume–region pair, with both quantities expressed as fractions.
Supplementary overlap measure that penalizes false positives and false negatives symmetrically; stricter than Dice, hence separates methods whose Dice values are compressed near the upper end of the scale.
Monotone transform of Dice per pair and therefore largely redundant; shares the insensitivity of Dice to error location; absolute values are not comparable with Dice-based results reported elsewhere.
Precision. The reliability of predicted FG voxels is given by ( Taha and Hanbury, 2015 ) Precision(Mr,Mr)=TP+FPTP. (13) It is equivalently the positive predictive value.
Proportion of predicted FG voxels that are correct; isolates the FP component that Dice merges with FN , thereby quantifying over-segmentation.
Driven close to its maximum by conservative predictions, hence uninformative unless read jointly with recall; degenerate on target-absent pairs; disregards the spatial distribution of false positives.
Recall. The recovery of reference FG voxels is ( Taha and Hanbury, 2015 ) Recall(Mr,Mr)=TP+FNTP. (14) It is equivalently the sensitivity or true-positive rate.
Proportion of reference FG voxels recovered by the prediction; quantifies under-segmentation, the error type with the greatest clinical cost in surgical and radiotherapy planning.
Attains its maximum for any model labeling the entire brain as FG; entirely insensitive to false positives; degenerate on target-absent pairs.
Specificity. The correct rejection of BG voxels is ( Taha and Hanbury, 2015 ) Specificity(Mr,Mr)=TN+FPTN. (15) It is equivalently the true-negative rate.
Proportion of BG voxels correctly rejected as BG; describes the negative class, which the overlap metrics exclude; provides a volume-wide check against indiscriminate FG activation.
Saturates near 100% because BG voxels dominate the volume, hence offers little resolution among models; a high value may still correspond to a clinically significant number of false-positive voxels.
Table 2: Definitions, main advantages, and main limitations of the volume-level evaluation metrics.
Figure 4: Training and validation dynamics of the compared models on BraTS20.
WT
TC
ET
Model
Dice
HD95
Dice
HD95
Dice
HD95
ALPNet ( Ouyang et al., 2020 )
76.38±6.39
41.48±3.86
56.74±10.96
47.80±18.76
47.88±4.35
36.36±23.24
PANet ( Wang et al., 2019 )
80.44±5.38
32.48±2.46
75.33±2.06
26.14±7.86
62.65±2.78
15.65±2.64
Q-Net ( Shen et al., 2023 )
82.74±0.58
37.23±1.53
66.53±10.30
45.36±4.22
55.06±3.23
40.43±3.15
CATNet ( Lin et al., 2023b )
82.48±0.10
37.70±0.98
68.89±2.60
39.01±14.16
69.34±1.17
18.08±5.25
GMRD ( Cheng et al., 2024 )
85.98±1.05
26.13±3.64
79.09±1.20
24.16±5.95
68.95±7.75
15.57±9.13
Table 3: BraTS20 few-shot results reported as mean ± sample standard deviation over five deterministic support draws. Best point estimates are bold.
WT
TC
ET
Model
Dice
HD95
Dice
HD95
Dice
HD95
U-Net ( Ronneberger et al., 2015 )
86.20
28.01
80.20
25.78
75.72
17.65
UNet++ ( Zhou et al., 2019 )
86.35
27.61
80.61
27.68
74.77
23.74
TransUNet ( Chen et al., 2021 )
82.72
40.91
75.24
39.16
70.43
16.79
SwinUNet ( Cao et al., 2022 )
84.29
24.24
74.01
23.28
68.23
19.07
VMUNet ( Ruan et al., 2025 )
88.32
18.41
80.77
20.97
75.09
13.44
Table 4: BraTS20 volume-level Dice and HD95 for non-FSS baselines and FSS-UBrain. The best point estimate in each column is shown in bold.
BraTS-Africa
BraTS23
Model
Mean Dice
Mean HD95
Mean Dice
Mean HD95
ALPNet ( Ouyang et al., 2020 )
66.83±0.85
40.40±2.05
69.15±4.32
40.50±4.10
PANet ( Wang et al., 2019 )
74.37±1.08
16.29±1.01
76.77±0.28
21.94±1.17
Q-Net ( Shen et al., 2023 )
68.83±6.04
28.94±0.43
71.02±0.76
42.39±0.02
CATNet ( Lin et al., 2023b )
70.65±2.01
26.56±1.95
71.34±0.43
36.21±2.61
GMRD ( Cheng et al., 2024 )
77.65±0.86
18.46±1.14
81.84±0.98
13.44±1.01
Table 5: Target-supported cross-cohort mean results. BraTS-Africa values are mean ± sample standard deviation over five deterministic support draws; BraTS23 values are mean ± sample standard deviation over two deterministic support draws (seeds 1337 and 1338). Best point estimates are bold.
Source support
Target support
Model
Mean Dice
Mean HD95
Mean Dice
Mean HD95
ALPNet ( Ouyang et al., 2020 )
61.32±2.82
38.07±11.75
66.83±0.85
40.40±2.05
PANet ( Wang et al., 2019 )
73.05±1.71
16.90±0.77
74.37±1.08
16.29±1.01
Q-Net ( Shen et al., 2023 )
70.60±2.49
29.45±1.29
68.83±6.04
28.94±0.43
CATNet ( Lin et al., 2023b )
72.59±1.21
25.82±3.46
70.65±2.01
26.56±1.95
GMRD ( Cheng et al., 2024 )
76.62±2.18
18.33±2.43
77.65±0.86
18.46±1.14
Table 6: BraTS-Africa mean results under source and target support (mean ± sample standard deviation over five draws; same 85 query cases). Best point estimates within each support condition are bold.
Metric
Full model
w/o DMP
w/o SGB
w/o PC
w/o BR
Mean Dice
83.13
82.33
82.42
82.75
82.98
Mean HD95
8.20
9.27
9.96
9.86
9.59
Table 7: BraTS20 mean component-ablation results across WT, TC, and ET. Best point estimates are bold.
Figure 5: Qualitative few-shot comparisons using four-modality MRI input; only FLAIR is displayed. SP denotes support, GT ground truth, and Pred prediction.
Cohort
Train Hard Neg
Train Full Neg
Val Hard Neg
Val Full Neg
Mean Dice
Mean HD95
Mean IoU
Precision
Recall
Specificity
FP-neg-rate
BraTS20
0
0
0
0
75.46
41.97
63.93
72.82
83.33
99.89
0.35
20
0
25
0
82.72
10.68
73.91
82.11
86.99
99.93
0.16
0
10
0
15
81.52
12.81
72.41
80.68
86.57
99.92
0.23
10
5
15
10
83.21
11.42
74.27
82.09
87.06
99.93
0.15
20
10
25
15
83.13
8.20
74.34
82.59
86.81
99.93
0.10
30
15
35
20
84.19
7.61
75.90
83.75
87.45
99.94
0.12
Table 8: BraTS20 sensitivity to hard- and fully negative-query sampling. Mean Dice and mean HD95 are unweighted averages across WT, TC, and ET; ratios denote training hard/full and validation hard/full sampling. Best point estimates are bold.
Model
Params (M)
FLOPs (G)
MACs (G)
Without DMP
19.83
200.45
100.22
Without SGB
41.82
351.05
175.53
Without PC
36.88
328.39
164.19
Without BR
42.61
347.00
173.50
FSS-UBrain (Full model)
42.65
352.05
176.03
Table 9: Model-side complexity of FSS-UBrain ablations per 4×256×256 one-shot support–query episode at batch size one. Test-time augmentation and volume post-processing are excluded. Lowest values are bold.
Model
Params (M)
FLOPs (G)
MACs (G)
SwinUNet ( Cao et al., 2022 )
6.94
6.90
3.45
U-Net ( Ronneberger et al., 2015 )
7.70
73.91
36.96
UNet++ ( Zhou et al., 2019 )
36.62
276.83
138.42
TransUNet ( Chen et al., 2021 )
105.32
77.09
38.55
VMUNet ( Ruan et al., 2025 )
44.28
19.66
9.82
SAM ( Kirillov et al., 2023 )
93.74
976.50
488.25
Table 10: Model-side complexity at batch size one. Support-conditioned methods are profiled per support–query episode and non-FSS methods per query slice; test-time augmentation and volume post-processing are excluded. Lowest values are bold separately for the non-FSS group (upper block, including SAM) and the support-conditioned group (lower block, including FSS-UBrain).
Few-shot brain tumor segmentation remains challenging due to noisy support masks, inter-patient variations between support and query images, and the lack of pixel-wise confidence estimation. This study proposes RUFNet, a Hybrid Mamba-based few-shot framework that combines support mask refinement with uncertainty-aware posterior fusion. To preserve support-query dependencies with manageable cost, RUFNet adopts a Hybrid Mamba interaction backbone with linear complexity. To reduce support-mask noise, an Attention-Guided Mask Refinement module (AGMR) uses query features to recalibrate support masks and improve prototype consistency. To handle ambiguous predictions, an Uncertainty-Aware Posterior Fusion module (UAPF) estimates pixel-wise variance and adaptively balances few-shot predictions with query-aligned priors. On the Brain Tumor Segmentation Challenge (BraTS) 2020 dataset, RUFNet achieves Dice coefficients of 84.3% and 86.1% in the 1-way 1-shot and 1-way 5-shot settings, respectively, outperforming the compared state-of-the-art methods. These results suggest that Hybrid Mamba interaction, mask refinement and uncertainty modelling can improve the robustness of few-shot medical image segmentation. The official implementation code is available at https://github.com/hdy6438/RUFNet.
Dongyi He, Xiangkai Wang, Binbing Xu +5
Department of Language Science and Technology, The Hong Kong Polytechnic University, Hong Kong SAR, China · School of Artificial Intelligence, Chongqing University of Technology, Chongqing, China · Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang, China +1
Accurate segmentation of brain tumour sub-regions from multi-parametric MRI is critical for treatment planning yet remains challenging due to morphological variability, class imbalance, and overlapping appearances of tumour regions across imaging sequences. We propose SegGuidedNet, a three-dimensional residual encoder--decoder network introducing a novel SegAttentionGate module that explicitly supervises the decoder to produce spatially discriminative attention maps for each tumour sub-region necrotic core, peritumoral oedema, and enhancing tumour via a lightweight auxiliary loss, adding less than 0.2% parameter overhead. This sub-region supervision maintains decoder discriminability between visually ambiguous classes while providing free-of-cost spatial interpretability at inference without any post-hoc explanation method. Evaluated independently on BraTS2021 and BraTS2023 GLI across 251 held-out subjects each, SegGuidedNet achieves mean Dice of 0.905 (ET= 0.873, TC=0.906, WT=0.935) and 0.897 (ET=0.859, TC=0.902, WT=0.931) respectively, surpassing ensemble-based nnU-Net and HNF-Netv2 as a single model and approaching Swin UNETR a 10-model ensemble within 2--4 Dice points at a fraction of the inference cost. The consistency of results across two benchmark editions further confirms the generalisability of the proposed approach, offering competitive accuracy with built-in interpretability in a lightweight, clinically practical framework.
Hasaan Maqsood, Saif Ur Rehman Khan, Sebastian Vollmer +2
National University of Sciences and Technology (NUST), Islamabad, Pakistan. · Department of Computer Science, Rhineland-Palatinate Technical University of Kaiserslautern-Landau, Kaiserslautern, 67663, Germany. · German Research Center for Artificial Intelligence, Kaiserslautern, 67663, Germany. +2
Existing brain tumor segmentation methods usually utilize multiple Magnetic Resonance Imaging (MRI) modalities in brain tumor images for segmentation, which can achieve better segmentation performance. However, in clinical applications, some modalities are often missing due to resource constraints, resulting in significant performance degradation for methods that rely on complete modality segmentation. In this paper, we propose a Multimodal feature distillation with Mamba-Transformer hybrid network (MMTSeg) for accurate brain tumor segmentation with missing modalities. We first employ a Multimodal Feature Distillation (MFD) module to distill feature-level multimodal knowledge into different unimodalities to extract complete modality information. We further develop an Unimodal Feature Enhancement (UFE) module to model the semantic relationship between global and local information. Finally, we built a Cross-Modal Fusion (CMF) module to explicitly align the global correlations across modalities, even when some modalities are missing. Complementary features within and across modalities are refined by the Mamba-Transformer hybrid architectures in both the UFE and CMF modules, dynamically capturing long-range dependencies and global semantic information for complex spatial contexts. A boundary-wise loss function is employed as the segmentation loss of the proposed MMTSeg to minimize boundary discrepancies for a distance-based metric. Our ablation study demonstrates the importance of the proposed feature enhancement and fusion modules in the proposed network and the Transformer with Mamba block for improving the performance of brain tumor segmentation with missing modalities. Extensive experiments on the BraTS 2018 and BraTS 2020 datasets demonstrate that the proposed MMTSeg framework outperforms state-of-the-art methods when modalities are missing.
Ming Kang, Fung Fung Ting, Shier Nee Saw +3
School of Information Technology, Monash University, Malaysia Campus, Subang Jaya 47500, Malaysia · Augmented Intelligence and Multimodal analytics (AIM) for Health Lab, Faculty of Information Technology, Monash University, Clayton VIC 3800, Australia · Monash Medical AI Group, Monash University, Clayton VIC 3800, Australia +1