An nnU-Net-based framework with adaptive feature representation for 3D brain tumor segmentation

Introduction

Gliomas are common primary malignant tumors of the central nervous system, characterized by marked spatial variability and invasive growth. Multimodal magnetic resonance imaging (MRI) provides complementary information for tumor diagnosis, treatment planning, and response assessment. Accurate tumor segmentation is therefore essential for quantifying tumor burden and monitoring disease progression (1). However, manual slice-by-slice delineation is time-consuming and operator-dependent, making it difficult to meet the needs of large-scale clinical workflows. The brain tumor segmentation (BraTS) benchmark has promoted standardized evaluation of BraTS through multi-center data and expert annotations, commonly using whole tumor (WT), tumor core (TC), and enhancing tumor (ET) as evaluation regions (2).

Deep learning methods for medical image segmentation have largely been developed from U-shaped encoder-decoder architectures. U-Net has established a widely used framework by combining semantic information and spatial details through skip connections (3), and its three-dimensional (3D) extension further enables volumetric context modeling (4). Beyond architecture, segmentation performance also depends on preprocessing, training strategy, inference procedure, and post-processing. As a strong self-configuring baseline, nnU-Net provides a reproducible pipeline for evaluating architectural modifications under controlled settings (5). Because benchmark results are sensitive to implementation details, rigorous experimental control and transparent evaluation protocols are also necessary (6).

Despite these advances, BraTS remains challenging. Tumor subregion boundaries often show low contrast and irregular shapes, making overlap-based optimization insufficient for boundary consistency. ET regions are usually small and highly imbalanced, and may be overlooked during downsampling and deep feature compression. In addition, 3D networks often require strong feature compression because of memory constraints, which may increase channel redundancy and weaken fine-grained feature selection.

Several strategies have been proposed to address these issues. Boundary loss provides distance-field-based supervision to improve boundary consistency (7). Squeeze-and-excitation (SE) attention enhances discriminative channel responses through channel recalibration (8), while CondConv improves sample-adaptive representation by combining input-conditioned expert convolution kernels (9). However, these techniques are often introduced as isolated modules, and their complementary roles within a strong self-configuring baseline have not been sufficiently examined.

To support cross-modality evaluation, KiTS2019 provides a public benchmark for kidney and kidney tumor segmentation in contrast-enhanced computed tomography (CT), and it was used as the second dataset in this study (10). Recent advances in medical image segmentation further provide the methodological context for the proposed framework. Medical image segmentation has largely developed from U-shaped encoder-decoder architectures. U-Net, by integrating shallow details and deep semantics through skip connections, established the fundamental framework for subsequent methods (3). The 3D U-Net model extends this structure to volumetric data and enables the network to directly model 3D spatial context (4). In such frameworks, mechanisms such as attention gating can enhance the selectivity of target regions and improve the segmentation performance of small structures or low-contrast regions (11).

In addition to network structure, medical segmentation performance depends on preprocessing, training strategy, inference procedure, and post-processing. nnU-Net unifies these key steps through a dataset-adaptive self-configuration strategy, providing a highly reproducible baseline to evaluate whether structural changes bring real benefits under a unified pipeline (5). Benchmark performance is also sensitive to implementation details, which further highlights the need for rigorous experimental control and transparent evaluation protocols (6).

In recent years, Transformers have been introduced into medical segmentation to enhance long-range dependency modeling capabilities. TransUNet integrates a Transformer encoder with a U-Net-style decoder and introduces a global self-attention mechanism (12). UNETR further explores the use of Transformers as 3D encoders for the segmentation of volume data (13). Furthermore, hierarchical visual Transformers, such as Swin Transformer-based models, achieve efficient multi-scale modeling through windowed self-attention (14,15). These models have been incorporated into U-shaped segmentation networks. While they offer advantages in global relation modeling, they are frequently accompanied by elevated computational and memory overhead in 3D scenarios. Moreover, achieving stable benefits in boundary quality and small-volume targets often requires more appropriate feature selection and supervision strategies.

In addition to these directions, lightweight module enhancement and boundary-aware supervision have also received attention. With regard to feature enhancement, the SE module improves channel selectivity through channel recalibration (8). The convolutional block attention module (CBAM) combines channel and spatial attention to suppress redundant responses (16). CondConv improves sample-adaptive representation by combining input-conditioned expert convolution kernels (9). At the supervision level, boundary loss uses distance-field information to provide direct constraints on the segmentation interface and can complement region-based losses in class-imbalanced or boundary-ambiguous scenarios (7).

Recent developments in medical image segmentation have further expanded beyond conventional convolutional neural network (CNN)- and transformer-based architectures. Foundation and generalist segmentation models, such as MedSAM (17) and SegVol (18), have explored promptable or universal segmentation across heterogeneous medical imaging tasks. Large-scale systems such as TotalSegmentator (19) have also demonstrated the potential of training segmentation models with broad anatomical coverage. Meanwhile, efficient 3D Transformer variants, such as UNETR++ (20) and Slim UNETR (21), have focused on improving the accuracy-efficiency trade-off in volumetric segmentation. More recently, Mamba-based and state-space-model-based architectures, such as Swin-UMamba (22), have been introduced to strengthen long-range dependency modeling with lower computational complexity. In addition, recent benchmarking studies such as nnU-Net Revisited (23) have emphasized that new 3D medical segmentation methods should be evaluated using strong baselines, transparent protocols, and computational-cost-aware comparisons.

Taken together, these studies show that current medical image segmentation research increasingly emphasizes generalization, volumetric efficiency, long-range dependency modeling, and rigorous validation. However, many recent models require large-scale pretraining, prompt-based interaction, or substantial architectural redesign. In contrast, this study focuses on a controlled, baseline-preserving enhancement within the nnU-Net v2 framework. Rather than replacing the segmentation backbone, conditional convolution and squeeze-and-excitation with boundary-aware learning network (CondSEB-Net) aims to improve adaptive feature representation, channel selectivity, and boundary-aware supervision while maintaining the reproducibility of the self-configuring nnU-Net pipeline.

In this study, we propose CondSEB-Net as a controlled enhancement of nnU-Net v2 rather than a replacement for the nnU-Net pipeline. The proposed design is task-driven: CondConv addresses inter-case morphological and textural heterogeneity, three-dimensional squeeze-and-excitation (3DSE) attention reduces multi-scale channel redundancy, and boundary-aware supervision constrains ambiguous boundaries in small or low-contrast regions. Thus, CondSEB-Net is not an arbitrary aggregation of existing modules, but a baseline-preserving design that links sample-level adaptation, channel-level recalibration, and contour-level supervision.

The main contributions of this study are as follows:

• We propose a controlled nnU-Net v2-based framework that integrates conditional convolution (CondConv), multi-level 3D channel attention, and boundary-aware supervision to improve adaptive feature representation and boundary consistency.
• We clarify the complementary roles of the three components in addressing inter-case heterogeneity, channel redundancy, and boundary ambiguity.
• We evaluate the proposed method on BraTS2020 and KiTS2019 (2,10) under a unified training and inference protocol and analyze its practical significance in terms of computational overhead, boundary quality, and small-region segmentation performance.

We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0792/rc).

Methods

Representative imaging modalities and annotations from BraTS2020 and KiTS2019 are shown in Figure 1.

Figure 1 Dataset and annotation examples. Top: BraTS2020 four-modal MRI input and annotations (WT/TC/ET) overlaid on FLAIR; bottom: KiTS2019 CT input and annotations (kidney/tumor) overlaid. Different colors represent different categories, and z indicates the axial slice number. BraTS, brain tumor segmentation; CT, computed tomography; ET, enhancing tumor; FLAIR, fluid-attenuated inversion recovery; GT, ground truth; KiTS, kidney tumor segmentation; MRI, magnetic resonance imaging; T1, T1-weighted imaging; T2, T2-weighted imaging; T1ce, T1-weighted contrast-enhanced; TC, tumor core; WT, whole tumor.

Overall architecture

As illustrated in Figure 2, to better address feature redundancy and inter-case variability in medical image segmentation, dynamic convolution (CondConv) and SE attention modules are incorporated into the nnU-Net backbone to enhance adaptive feature representation. The proposed CondSEB-Net is built upon the nnU-Net v2 framework, inheriting its self-configuring training paradigm and 3D encoder-decoder architecture. Specifically, the model retains key design elements, including multi-scale feature extraction, skip connections, and deep supervision, ensuring reproducibility and fair comparison with baseline methods. On this basis, the proposed network is further enhanced at both the architectural and supervisory levels, leading to improved segmentation performance, particularly in challenging tumor boundary regions.

Figure 2 A schematic diagram of the overall structure of CondSEB-Net. The method uses the 3D encoder-decoder architecture of nnU-Net v2 as its backbone, retaining designs such as multi-scale feature extraction, skip connections, and deep supervision. SE channel attention is introduced at each scale stage for channel recalibration; the remaining structure and training/inference process remain unchanged to ensure fairness and reproducibility in the control experiments. 3D, three-dimensional; SE, squeeze-and-excitation.

At the structural level, all 3×3×3 convolutions in the network are implemented using CondConv3D dynamic convolutions (see the blue convolution blocks in the figure) to enhance the adaptive representation of morphological differences and heterogeneous textures in different cases. Concurrently, multi-level SE channel attention is inserted at each scale stage to recalibrate the channels of multi-scale features. This process highlights key semantic channels and suppresses redundant responses.

At the supervisory level, a Boundary loss is introduced on top of the default Dice + cross-entropy (CE) loss in nnU-Net v2. Distance field constraints further strengthen boundary alignment and reduce misclassification and adhesion at complex edges. All these modifications were implemented without altering the core topology of nnU-Net v2 or the self-configured training/inference process. This ensured that rigorous and fair ablation experiments and baseline comparisons could be facilitated. The specific implementation and training settings of each module will be detailed as follows.

Improved network architecture

This section delineates the network modifications and embedding methods of CondSEB-Net, which are based on the nnU-Net v2 backbone structure. To ensure fairness and reproducibility, with the exception of the structural modules and loss terms introduced in this study, the data preprocessing, self-configured network parameters, inference, and post-processing follow the nnU-Net v2 pipeline, while the stated optimizer and training schedule are applied consistently to both the nnU-Net v2 baseline and CondSEB-Net.

Backbone: nnU-Net v2 setup

This study uses nnU-Net v2 as the backbone network. Its core is a 3D U-shaped encoder-decoder structure: the encoder extracts multi-scale semantic features through progressive downsampling, and the decoder restores spatial resolution through progressive upsampling, fusing with skip connections at the corresponding scale to preserve detail information, thus outputting voxel-level segmentation results.

nnU-Net v2 employs a dataset-adaptive self-configuration strategy, automatically determining key network and training parameters (such as the number of network scales, number of channels per layer, convolution stacking depth, patch size, and batch size) based on the voxel spacing, foreground scale, and memory budget and provides a strongly reproducible baseline on different datasets. During training, 3D patch training and default data augmentation strategies are used, along with deep supervision to improve multi-scale learning performance; during inference, sliding window prediction is used, and overlapping regions are fused to obtain complete volumetric segmentation results. Except for the changes introduced in the following sections, all other settings are consistent with the nnU-Net v2 baseline.

CondConv integration

To enhance the network’s capacity for adaptive representation of morphological differences and heterogeneous textures across different cases, this paper introduces CondConv into the convolutional module of nnU-Net v2. CondConv, as proposed by Yang et al., generates weight coefficients through routing/gating branches driven by input features and weights multiple “expert kernels” to construct corresponding equivalent convolution parameters for each sample. This achieves dynamic parameterization of the convolution operator and adaptive feature extraction.

In implementation, the 3×3×3 standard 3D convolutions (Conv3d) in each scale’s stacked conv blocks of the encoder and decoder are uniformly replaced with CondConv3D, while the remaining structures (downsampling/upsampling, skip connections, and multi-scale feature fusion) are left unchanged. Convolutions of the 1×1×1 type (for example, channel transformation and segmentation heads) are regarded as standard convolutions, with the purpose of maintaining training stability and control of additional overhead.

CondConv retains K expert convolutional kernels {Wk}k=1K (and the corresponding biases {b_k}) for each convolutional layer; this study uses K=4 by default. The routing branch generates a gating weight vector α(x)∊(0,1)^K for each input sample, which is generated using a combination of global average pooling and 1×1×1 convolution, and outputs the gating coefficients via a Sigmoid function:

α(x)=σ(Conv1×1×1(GAP(x)))[1]

In this model, GAP(∙) represents adaptive global average pooling for the spatial dimension, and σ(∙) is the Sigmoid function. Therefore, given Sigmoid’s role in generating the gating coefficients, there is no necessity for the expert weights to be normalized. Using Sigmoid facilitates the concurrent activation of multiple expert kernels (non-competitively), a configuration that has been empirically substantiated to engender enhanced stability in heterogeneous 3D scenes. Subsequently, the expert kernels are weighted and aggregated according to the sample to obtain the equivalent convolution parameters for that sample:

W(x)=∑k=1Kαk(x)Wk,b(x)=∑k=1Kαk(x)bk[2]

In this context, the letter α_k(x) denotes the Kth component of the variable α(x). The variables W(x) and b(x) are then used to perform the 3D convolution operation. To enhance computational efficiency, the implementation of dynamic convolution computation is achieved through the use of batch-dimensional parallel grouped convolution. The overall structure of the 3D CondConv module is illustrated in Figure 3.

Figure 3 3D CondConv module schematic diagram. The input feature X first generates sample-related mixing coefficients through a routing branch, and the K groups of expert convolution kernels are weighted and aggregated to obtain equivalent convolution parameters. The aggregated parameters are then used for CondConv3D, followed by normalization and nonlinear mapping. 3D, three-dimensional; CondConv, conditional convolution; FC, fully connected layer; GAP, global average pooling.

Multi-level 3DSE channel attention integration

In accordance with the research findings of previous channel-attention studies (8,16), it can be concluded that channel attention has the capacity to explicitly model inter-channel dependencies. This, in turn, has the effect of improving feature selectivity and suppressing redundant responses. To enhance the network’s ability to select discriminative channels under multi-scale representation, this study introduces a 3DSE channel attention module into the multi-scale structure of nnU-Net v2. The 3DSE module performs global aggregation of the 3D spatial dimensions (H×W×D) of the volume data in the “squeeze” stage, in contrast to the 2DSE module, which makes it more suitable for 3D segmentation tasks. This study performs multi-level integration at the “stage” granularity: in the encoder, 3DSE is inserted at the end of the stacked conv blocks of each stage for the output features of that stage; in the decoder, 3DSE is inserted at the end of the feature fusion convolution blocks of each stage to recalibrate the channels of the fused features. It is important to note that this design does not modify the normalization and activation sequence within the original convolutional units of nnU-Net v2. However, it does perform channel recalibration at the final output of each stage.

The 3DSE computation follows the standard squeeze-excitation process: the initial stage of the procedure is the subjection of input feature x∈ℝB×C×H×W×D to global average pooling in 3D space (H×W×D). The purpose of this process is to obtain the channel description vector z∈ℝB×C:

zb,c=1HWD∑h=1H∑w=1W∑d=1Dxb,c,h,w,d[3]

Subsequently, a channel weight vector s∈(0,1)B×C is generated through two fully connected layers, and the input features are recalibrated by channels: x′=s⊙x is represented by the symbol ⊙, which denotes the process of multiplication across channels. The default value of r is set to 16, a choice that allows for the effective management of the number of additional parameters, while ensuring the maintainability of the module’s lightweight nature. During the fusion process with skip connections, at each scale, the decoder first concatenates the upsampled features from the previous level with the corresponding skip features from the encoder in the channel di-mension, and then obtains the fused features through a convolutional fusion block. Subsequently, 3DSE is introduced at the end of this fusion block to recalibrate the fused features by channels. This process further highlights effective channel responses related to the target region and suppresses the propagation of redun-dant features. Figure 4 presents a schematic representation of the 3DSE module.

Figure 4 3DSE channel attention module diagram. The input features are subjected to global average pooling in the spatial dimension (H×W×D) to obtain channel description vectors, and channel weights are generated through two layers of MLP to recalibrate the features and enhance the discriminative response. 3DSE, three-dimensional squeeze-and-excitation; MLP, multi-layer perceptron.

Boundary-aware objective function

Kervadec et al. (7) proposed a novel loss function that is aware of boundaries. This function explicitly characterises the geometric information of the target contour using a signed distance map (SDM). This allows the optimization process to directly address segmentation boundary errors, thus improving boundary consistency. To optimise both region overlap and boundary alignment simultaneously, this paper employs a composite objective function that is aware of the boundaries. This function consists of Dice loss, CE loss and boundary loss. The overall formulation of the boundary-aware composite loss is illustrated in Figure 5.

L=LDice+LCE+λLBoundary[4]

Figure 5 Schematic diagram of the boundary-aware composite loss function. The model output logits are processed by the softmax function to obtain a multi-class probability map P, while the signed distance map, SDM(GT), is calculated from the GT. The boundary loss uses P and SDM(GT) to impose direct constraints on the segmentation interface and strengthen boundary alignment. CE, cross-entropy; GT, ground truth; SDM, signed distance map.

In this context, λ denotes the weight of the boundary term. In this study, λ=0.5 is selected on the validation set to balance the contributions of region supervision and boundary supervision.

In the context of multi-class segmentation, the network output logits are processed by the softmax function to yield the probability map P. To streamline the notation, the batch dimension is omitted below and is represented by P∈ℝC×H×W×D. The implementation and aggregation strategy of nnU-Net v2 is adopted by both dice loss and CE loss, in order to constrain region overlap. Boundary loss is then added to enhance boundary alignment.

SDM

The use of SDM is intended to explicitly encode the geometric information of the ground truth boundaries. Given a ground truth segmentation mask G, a binary mask G_c is constructed for each target category c. Its complement (background) is defined as Gc¯. This paper adopts the symbol convention of “negative inside the mask, positive outside, and 0 at the boundary”, defining the SDM as follows:

ϕ(Gc)=d(Gc¯)−d(Gc)[5]

In this context, d(∙) denotes the Euclidean distance transformation, while G_c and Gc¯ represent the foreground and background regions, respectively. To maintain a stable numerical range, ϕ(∙) can undergo either normalization or truncation.

Boundary loss

Boundary loss employs the use of a SDM to weight errors in proximity to the boundary, making the training process more sensitive to interface misalignment. The study employs a boundary term in the form of a distance field to couple the predicted probability with SDM:

LBoundary=1|Ω|∑c∈Cfg∑x∈ΩPc(x)ϕ(Gc)(x)[6]

In this study, the set of voxels in the current training patch is denoted by Ω, the set of foreground classes by Ϲ_fg, and the predicted probability that voxel x belongs to class c by P_c(x). This term encourages the predicted boundary to align with the true boundary, which can improve the segmentation quality of thin structures and irregular contours (7).

Training and inference protocol

Cross-validation strategy

To obtain robust and reproducible evaluation results, this paper employs 5-fold cross-validation on both the BraTS2020 and KiTS2019 datasets. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The process of splitting is conducted at the subject level, which signifies that 3D volumetric data and their corresponding annotations for a particular case are consistently allocated to the same fold. This approach is used to prevent the occurrence of data leakage between the training and validation sets. In each cross-validation iteration, four folds are used for training and the remaining fold is used for validation. Subsequent to five training-validation iterations, the results from each fold are averaged to obtain the final result. To ensure a fair comparison of different methods under the same data split, all experiments in this paper use a fixed 5-fold splitting scheme (generated by the default mechanism of nnU-Net v2 and maintained unchanged), which is also consistent in subsequent ablation and control experiments.

Training settings

The training process is based on the standard 3D patch segmentation pipeline of nnU-Net v2, and uses dataset-specific automatically generated plans to determine basic configurations such as resampling spacing, patch size, batch size, and data augmentation. To ensure the comparability between the baseline and the proposed method, the training settings on BraTS2020 and KiTS2019 are consistent with the nnU-Net v2 baseline, with the exception of the modules and loss terms introduced in this paper. The optimiser under consideration is AdamW, with an initial learning rate of 1×10⁻², weight decay of 3×10⁻⁵, and learning rate decay according to the default strategy of nnU-Net v2. The training is scheduled to last for 200 epochs; each epoch comprises 250 training iterations and 50 validation iterations. To enhance the learning of small-volume foreground structures, a foreground oversampling strategy is adopted, with oversample_foreground_percent =0.33, and probabilistic oversampling is disabled. Furthermore, the implementation of deep supervision has been demonstrated to enhance the stability of multi-scale feature learning and training. Except for these settings, all other implementation details are consistent with nnU-Net v2.

Inference and post-processing

The inference phase follows the default nnU-Net v2 inference flow, employing a sliding window approach to divide the entire 3D volumetric data into blocks for prediction. The dimensions of the window are determined by dataset-specific plans (which are usually consistent with the patch settings during training), and an overlap ratio of 0.5 is set between adjacent windows. The weighted fusion process is executed in accordance with the default nnU-Net v2 strategy, to obtain the full-volume class probability map. Subsequently, voxel-level decisions are made on the probability map to generate the final segmentation labels. Post-processing also follows the default configuration of nnU-Net v2: if connected component analysis-based filtering is enabled for the corresponding task, the largest connected component is retained to remove small artifacts; otherwise, no additional post-processing is applied, and the network output is used directly for evaluation and statistics.

Results

Dataset and evaluation metrics

Dataset

The present study conducted experiments on two publicly available datasets from Medical Image Computing and Computer Assisted Intervention (MICCAI): BraTS2020 and KiTS2019. BraTS2020 provides preoperative multimodal 3D MRI images [T1-weighted imaging (T1), T1-weighted contrast-enhanced (T1ce), T2-weighted imaging (T2), and fluid-attenuated inversion recovery (FLAIR)] in NIfTI format, which have been published and preprocessed using a series of standardised registration, cranial removal, and resampling techniques. This study used 369 cases from the BraTS2020 training set that had been labelled for the purposes of model training and evaluation, employing 5-fold cross-validation at the subject level. The BraTS2020 dataset was obtained from the official BraTS2020 resource (https://www.med.upenn.edu/cbica/brats2020/data.html) and cited according to the original BraTS benchmark publication (2). To ensure fairness in comparisons between different methods, all comparative experiments used the same fixed 5-fold partitioning (generated and maintained by the default mechanism of nnU-Net v2), thus avoiding bias introduced by differences in data partitioning. Examples of the datasets and their corresponding annotations are shown in Figure 1.

BraTS2020’s voxel-level annotations encompass three distinct tumor subregions: necrotic/non-enhancing tumor core (NCR/NET, label 1), peritumoral edema (ED, label 2), and ET (label 4). During the training phase, the labels are remapped to {0,1, 2, 3} (background, NCR/NET, ED, ET) in order to adapt to the multi-class segmentation settings of nnU-Net v2. During the evaluation phase, three composite regions are constructed according to the general BraTS definition: the whole tumor (WT) is defined as {1, 2, 4}, the tumor core (TC) as {1, 4}, and the ET as {4}. The evaluation metrics are calculated for each of the WT/TC/ET composite regions.

KiTS2019 provides 3D volumetric data and voxel-level annotations for contrast-enhanced abdominal CT scans, with the task being kidney and renal tumor segmentation. The KiTS2019 dataset was obtained from the official KiTS19 challenge resource (https://kits19.grand-challenge.org/) and cited according to the original KiTS19 challenge publication (10). The annotations are constituted of three mutually exclusive labels: background [0], kidney [1], and tumor [2]. The study employs a case-level five-fold cross-validation setup, and in the evaluation phase, evaluation metrics are calculated independently for two foreground categories of kidney and tumor, respectively. The macro-average of the two foreground categories was used as the overall performance reference.

Evaluation metrics

The evaluation metrics used in this study included the Dice similarity coefficient (DSC), the 95% Hausdorff distance (HD95, mm), and the average symmetric surface distance (ASSD, mm). For any evaluation region (WT/TC/ET) or foreground category, let the predicted segment be denoted by P and the true value be denoted by G. The DSC is then defined as follows:

DSC(P,G)=2|P∩G||P|+|G|[7]

In this context, |∙| denotes the cardinality of the voxel set. The DSC is chiefly indicative of the consistency of region overlap and is highly sensitive to the adequacy of overall coverage; it is one of the most commonly used overlap indicators in medical segmentation. From a statistical perspective, DSC can also be expressed as follows:

DSC=2TP2TP+FP+FN[8]

In this study, TP, FP, and FN denote the number of true positives, false positives, and false negatives, respectively.

HD95 characterizes boundary errors. Surface point sets S(P) and S(G) are defined as the sets of surface points corresponding to the prediction and the ground truth values, respectively. It should be noted that these values are typically obtained by extracting surface voxels using a 3D binary mask. The minimum distance from a point to the set is defined as:

d(x,S)=miny∈s‖x−y‖2[9]

Based on this distance definition, the one-way surface distance from S(P) to S(G) and the reverse set can be obtained. The HD95 metric is derived by taking the 95th percentile of two one-way distance distributions and utilising the larger of the two values as the final metric. This approach mitigates the influence of a small number of outliers, such as isolated misclassifications, on the Hausdorff distance. All metrics presented in this study are derived from the full-volume inference results and subjected to uniform statistical analysis using the evaluation implementation of nnU-Net v2. This ensures the consistency and reproducibility of the evaluation criteria.

In addition to HD95, the ASSD was included as an additional boundary-specific metric to more comprehensively evaluate contour agreement. For the predicted segmentation surface S(P) and the ground-truth surface S(G), ASSD was calculated as the average of bidirectional surface distances between S(P) and S(G). Compared with HD95, which reflects relatively large boundary deviations, ASSD measures the average surface discrepancy and therefore provides complementary information on overall contour alignment. A lower ASSD indicates better boundary agreement.

Statistical analysis

All quantitative results were summarized as mean ± standard deviation (SD) across the five cross-validation folds. Because nnU-Net v2 and CondSEB-Net were evaluated using the same fixed five-fold data partition, paired statistical comparisons were performed based on fold-wise paired differences. For each metric, the mean paired difference was calculated as CondSEB-Net minus nnU-Net v2. For Dice-based metrics, a positive difference indicates improvement, whereas for distance-based metrics, including HD95 and ASSD, a negative difference indicates improvement. The 95% confidence intervals (CIs) of the paired differences were calculated using the t distribution, and two-sided paired t-tests were used for hypothesis testing. A P value <0.05 was considered statistically significant.

Implementation details

All experiments were performed on a single NVIDIA Tesla V100-SXM2 graphics processing unit (GPU), using PyTorch 2.5.1 and CUDA 11.8. This study uses the open-source nnU-Net v2 implementation for data preprocessing, training, and inference. The nnU-Net v2 implementation was obtained from the publicly available repository and cited according to the original nnU-Net publication (5). Available at: https://github.com/DIAGNijmegen/nnUNet_v2. Its dataset-specific self-configuration plans were used to automatically determine key settings such as resampling, normalization, and network input. The training of hyperparameters and optimization details are described in Section 2.4.2 and will not be repeated here.

In relation to the configuration of the dataset, BraTS2020 employs a 3D_fullres configuration, with a resampling interval of 1.0×1.0×1.0 mm. The four-modal inputs are then normalised using z-scores, and statistics are estimated within the foreground mask area according to the default nnU-Net v2 strategy. KiTS2019 employs the 3D_fullres configuration, with a resampling interval of 0.78126×0.78125×0.78125 mm. The CT intensity normalization process utilises nnU-Net v2’s CTNormalization function, and the foreground mask normalization function is not enabled.

During the inference stage, a sliding-window strategy was used to perform block-wise prediction on the entire volume, and the final probability map is generated according to the probability fusion strategy of overlapping regions. This paper fixes the window overlap ratio at overlap = 0.5. Unless stated otherwise, post-processing is conducted in accordance with the default settings of nnU-Net v2. All metrics are calculated using the built-in evaluation tool of nnU-Net v2 (BraTS2020 is statistically analysed on the WT/TC/ET composite region, and KiTS2019 is statistically analysed on the kidney and tumor categories). The results presented herein are derived from local five-fold cross-validation inference and evaluation, and have not been submitted to the official evaluation server.

Comparison with baselines

To verify the effectiveness of CondSEB-Net, this paper employs 3D volumetric segmentation models for comparative experiments. This ensures consistency between the input context and the inference process. The comparison methods included classic 3D CNNs (3D U-Net) and representative 3D Transformer segmentation models (UNETR, Swin-UNETR), with nnU-Net v2 serving as a robust baseline. All methods are trained and inferred under the same five-fold cross-validation partition, and a unified evaluation method is used to calculate Dice and HD95, ensuring fairness in the comparison.

The primary findings from the BraTS2020 experiments are outlined in Table 1, presenting the mean values and 5-fold outcomes. In summary, CondSEB-Net demonstrates superior Dice and lower HD95 in the three composite regions: WT, TC, and ET. In comparison with nnU-Net v2, the proposed model enhances the average Dice from 0.8525 to 0.8659, while concomitantly reducing the average HD95 from 10.57 to 9.08 mm. In the more challenging ET region (typically smaller in volume, with more severe class imbalance and more blurred boundaries), ours achieves an improved Dice of 0.8275 (0.8051→0.8275) and a reduced HD95 of 14.32 mm (16.19→14.32). This suggests that the proposed dynamic convolution and multi-level channel recalibration enhance the representation of heterogeneous textures and small-volume structures across cases, while boundary-aware supervision further improves spatial representation at complex edges.

To further assess whether the observed improvements were robust across cross-validation folds and whether the boundary improvement was not limited to HD95, we performed paired statistical comparisons between nnU-Net v2 and CondSEB-Net using the same five-fold data partition. As shown in Table 2, CondSEB-Net achieved higher Dice scores and lower HD95 and ASSD values than nnU-Net v2 across the reported BraTS2020 metrics. The average Dice increased from 0.8525±0.0217 to 0.8659±0.0184, with a mean paired difference of +0.0134 and a 95% CI of 0.0119 to 0.0148. The average HD95 decreased from 10.57±4.01 to 9.08±3.64 mm, with a mean paired difference of −1.49 mm and a 95% CI of −1.79 to −1.19 mm. In addition, the average ASSD decreased from 3.08±0.37 to 2.65±0.30 mm, with a mean paired difference of −0.43 mm and a 95% CI of −0.65 to −0.21 mm. These results indicate that CondSEB-Net achieved statistically supported improvements in both region overlap and boundary-related metrics compared with nnU-Net v2.

As illustrated in Table 3, CondSEB-Net also demonstrated superior or competitive performance on KiTS2019. Compared with nnU-Net v2, the proposed model improved the average Dice from 0.9096 to 0.9167 and achieved a consistent gain in the tumor category from 0.8412 to 0.8487. In addition, the average HD95 decreased from 6.33 to 6.06 mm. These results provide additional cross-dataset and cross-modality evidence, suggesting that CondSEB-Net has potential transferability across different segmentation tasks. However, this evaluation should not be interpreted as a substitute for external multi-center clinical validation under heterogeneous imaging protocols.

In summary, CondSEB-Net, while maintaining the strong baseline pipeline consistency of nnU-Net v2, can simultaneously improve both regional overlap and boundary localization accuracy, especially for small-volume targets with more complex boundaries (such as ET in BraTS and tumors in KiTS).

Ablation study

Component ablation

To evaluate the independent contributions of each module and their combined effect, this study uses nnU-Net v2 as a baseline. The following experiment was conducted in accordance with the same five-fold partitioning and consistent training settings as in section “Training and inference protocol”. The following techniques were gradually introduced: CondConv dynamic convolution, multi-level 3DSE channel attention, and boundary loss. The model that has been configured to enable all three simultaneously is designated as the final CondSEB-Net. The evaluation process employs the BraTS metric, utilising Dice and HD95 calculations on the WT/TC/ET composite region. As illustrated in Table 4, the five-fold mean results are reported. It is important to note that a higher mean value indicates a superior outcome, with a larger Dice value signifying a more favourable result and a smaller HD95 value indicating a more favourable outcome.

As demonstrated in Table 4, all three modules consistently yield performance gains at the baseline. In the single-module case, SE improves DSC from 0.8525 to 0.8568 (+0.0043) and reduces HD95 from 10.57 to 10.26 (−0.31). CondConv and Boundary loss also improve DSC and reduce HD95 (DSC 0.8563 and 0.8579, respectively, HD95 10.27 and 10.13, respectively). In the dual-module combination, performance is further enhanced, with a larger improvement in HD95 when combined with Boundary loss (e.g., SE + boundary loss: 9.58, CondConv + boundary loss: 9.62), suggesting that boundary supervision and feature adaptation selection are complementary and help improve boundary localization accuracy.

The model attains optimal overall performance when all three are enabled simultaneously (last row of Table 4). A comparison of the nnU-Net v2 baseline reveals an enhancement in the overall DSC from 0.8525 to 0.8659 (+0.0134). Additionally, the HD95 value experiences a decline from 10.57 to 9.08 mm (−1.49 mm). The findings suggest that CondConv assists in enhancing the adaptive representation of morphological and textural variations across cases. Furthermore, 3DSE enhances multi-scale feature selectivity through channel recalibration, while Boundary loss provides more direct boundary constraints. The combination of these three factors can lead to simultaneous improvements in region overlap and boundary localization accuracy, resulting in enhanced overall segmentation performance.

Region-wise analysis and discussion

In BraTS2020, the segmentation difficulty of different composite regions differs across tumor subregions. Typically, WT demonstrates a superior clarity of overall range, while TC/ET is more susceptible to boundary blurring, heterogeneous textures, and class imbalance. In particular, ET frequently exhibits small-volume and enhanced-edge irregularities, which result in relatively indirect constraints on boundary geometry in region overlap-based optimization. A combination of the results presented in Table 4 indicates that the three modifications have complementary performance improvements. CondConv enhances the adaptive representation of morphological and texture differences across cases through input-conditional dynamic kernel combinations, which is beneficial for maintaining stable local discrimination in complex heterogeneous regions. 3DSE performs channel recalibration on multi-scale features, which can strengthen effective channel responses related to tumors and suppress redundant features, thereby improving the separability of small-volume structures. Boundary loss provides more direct boundary supervision, making the optimization process more sensitive to interface position errors, thus leading to larger improvements in HD95. The combination of the three factors has been shown to improve DSC and further reduce HD95 (see Table 4). This finding indicates that the proposed method exhibits a consistent enhancement in both region overlap and boundary localization dimensions, rendering it more appropriate for managing target regions characterised by intricate boundaries and diminutive structures (e.g., ET).

Loss weight and sensitivity analysis

The weight coefficient, denoted by λ, of the boundary loss is utilised to balance the relative contributions of the region overlap term (Dice/CE) and the boundary constraint term in the optimization process. In the case of an overly diminutive value of λ, the impact of boundary supervision on training is found to be constrained; conversely, an excessively substantial value of λ has the potential to engender an overemphasization of errors in proximity to the interface, which may adversely affect regional consistency and overall convergence stability. In light of the validation set performance and training stability, this study adopts the default setting of λ=0.5. This value achieves an optimal trade-off between enhancing boundary localization accuracy (HD95) and preserving region overlap performance (DSC). It is imperative to note that all subsequent experiments are conducted with λ fixed at 0.5 to ensure fair and consistent comparisons between diverse methodologies.

In summary, this study, while maintaining consistent training and inference settings, verifies the complementarity and combined gain of each module through component ablation and weight setting analysis. The subsequent section will further demonstrate the performance differences of the model on complex boundaries and small-volume targets through visualisation comparison.

Qualitative results and case-level failure analysis

Qualitative comparison

To facilitate a more visual comparison of the differences in segmentation boundaries and details, this study selects three representative slices from the BraTS2020 cases used in the cross-validation experiments for the purpose of visual comparison (see Figure 6). These images showcase the segmentation contours of the input image (FLAIR/T1ce), the ground truth (GT) annotation, nnU-Net v2, and the proposed CondSEB-Net. A comprehensive evaluation reveals that CondSEB-Net exhibits consistent tumor coverage (WT/TC), while concurrently providing more stable characterisation of enhanced tumor regions (ET) and complex boundaries.

Figure 6 Qualitative comparison on BraTS2020. The proposed method shows more consistent boundaries, improved continuity in fragmented structures, and reduced spurious regions in challenging cases. BraTS, brain tumor segmentation; CondSEB-Net, conditional convolution and squeeze-and-excitation with boundary-aware learning network; ET, enhancing tumor; FLAIR, fluid-attenuated inversion recovery; GT, ground truth; T1ce, T1-weighted contrast-enhanced; TC, tumor core; WT, whole tumor.

In FLAIR, z=51, the lesion volume is substantial and the boundaries are irregular. A comparison of the predicted contours of CondSEB-Net with those of nnU-Net v2 reveals a higher degree of congruence with the GT at local edges. Furthermore, the ET/TC contours demonstrate a more continuous boundary morphology and reduce inconsistencies caused by local “jaggedness” and small-scale offsets. In FLAIR, z=104, the tumor structure exhibits multifocal/fragmented characteristics, representing a scenario with higher requirements for spatial continuity modelling. The visualisation results demonstrate that CondSEB-Net provides more consistent coverage of scattered regions and alleviates the fragmentation of some fine structures, thus presenting a more consistent region outline in complex morphologies. In the FLAIR model, with a z-value of 45, there is an increased probability of the production of isolated responses that are distant from the primary lesion. A comparison of CondSEB-Net’s prediction results with those of nnU-Net v2 reveals that the former exhibits more restrained responses in non-target regions, accompanied by a relative reduction in isolated small fragments (FPs). This finding suggests that CondSEB-Net possesses a stronger capacity to suppress irrelevant responses in complex backgrounds, which enhances the reliability of segmentation outcomes. The combination of the three sets of visualisation examples demonstrates that CondSEB-Net exhibits a more stable trend in terms of boundary consistency, coherence of fine structures, and suppression of non-target regions, particularly in regions such as ET that are small in size, have blurred boundaries, and are more susceptible to noise. This finding is consistent with the improvement trend of Dice/HD95 in the quantitative indicator of ET.

Case-level failure analysis

To provide a more balanced qualitative evaluation, we further examined representative challenging cases rather than only visually favorable examples. As shown in Figure 7, several residual error patterns remained. In the top panel, isolated false-positive responses were observed away from the main tumor region, indicating that tumor-like low-contrast signals may still lead to spurious predictions. In the middle panel, small separated ET components remained challenging, suggesting that tiny or spatially separated ET regions are still vulnerable to incomplete detection. In the bottom panel, a small satellite lesion and residual boundary discrepancy were observed, showing that boundary ambiguity in small or spatially separated regions is not fully resolved.

Figure 7 Representative challenging cases on BraTS2020. The top, middle, and bottom panels illustrate isolated false-positive responses, small separated enhancing tumor components, and residual boundary discrepancy, respectively. BraTS, brain tumor segmentation; CondSEB-Net, conditional convolution and squeeze-and-excitation with boundary-aware learning network; ET, enhancing tumor; FLAIR, fluid-attenuated inversion recovery; GT, ground truth; T1ce, T1-weighted contrast-enhanced; TC, tumor core; WT, whole tumor.

These observations indicate that CondSEB-Net improves boundary consistency and small-region segmentation, but does not completely eliminate FPs, tiny-lesion errors, or contour ambiguity. Future improvements may further consider lesion-size-aware learning, uncertainty estimation, and clinically guided post-processing to improve reliability in difficult cases.

Efficiency analysis

To evaluate the computational burden introduced by CondSEB-Net, we measured parameter count, floating point operations (FLOPs), peak inference GPU memory consumption, and inference latency under the same hardware and inference settings. Inference latency was calculated as the average processing time per case under the same sliding-window inference setting, excluding disk input/output. Peak GPU memory consumption was recorded during inference.

As shown in Table 5, CondSEB-Net introduced only a limited increase in computational cost compared with nnU-Net v2. On BraTS2020, the parameter count increased from 32.54 to 33.27 M, corresponding to a 2.24% increase. FLOPs increased from 522.67 to 531.84 G, corresponding to a 1.75% increase. Peak inference GPU memory consumption increased from 5.8 to 6.1 GB, corresponding to a 5.17% increase. Inference latency increased from 2.7 to 2.9 s/case, corresponding to a 7.41% increase.

On KiTS2019, the parameter count increased from 31.03 to 31.29 M, corresponding to a 0.84% increase. FLOPs increased from 512.42 to 517.63 G, corresponding to a 1.02% increase. Peak inference GPU memory consumption increased from 5.4 to 5.6 GB, corresponding to a 3.70% increase. Inference latency increased from 2.6 to 2.7 s/case, corresponding to a 3.85% increase.

These results indicate that CondSEB-Net introduced additional but limited computational overhead. The increased inference-time cost mainly comes from conditional kernel aggregation in CondConv and channel recalibration in the 3DSE module. In contrast, the boundary-aware loss only affects the training objective and does not introduce additional inference-time parameters or FLOPs. Therefore, the proposed method should be interpreted as an accuracy-cost trade-off rather than an accuracy-only improvement. Considering that CondSEB-Net improved the average DSC from 0.8525±0.0217 to 0.8659±0.0184 and reduced the average HD95 from 10.57±4.01 to 9.08±3.64 mm on BraTS2020, the additional overhead may be acceptable in scenarios where improved boundary delineation and small-region segmentation are prioritized.

Discussion

The improvement achieved by CondSEB-Net should be interpreted in the context of a strong nnU-Net v2 baseline. Because nnU-Net v2 already provides a highly optimized self-configuring pipeline, large absolute gains are difficult to obtain through architectural modification alone. Therefore, CondSEB-Net should be evaluated not only by DSC improvement, but also by fold-wise robustness, boundary localization, small-region sensitivity, and computational overhead. Unlike recent foundation, Transformer-based, or Mamba-based segmentation models, CondSEB-Net does not rely on large-scale pretraining, promptable segmentation, or substantial backbone redesign. Instead, it evaluates targeted modifications within a controlled and reproducible nnU-Net v2 setting (23).

On BraTS2020, CondSEB-Net improved the average DSC from 0.8525±0.0217 to 0.8659±0.0184, with a mean paired difference of +0.0134, a 95% CI of 0.0119 to 0.0148, and P<0.001. It also reduced the average HD95 from 10.57±4.01 to 9.08±3.64 mm and the average ASSD from 3.08±0.37 to 2.65±0.30 mm, suggesting improved boundary agreement. These statistically supported improvements were particularly meaningful for ET segmentation, where small lesion volume, class imbalance, and blurred boundaries remain challenging.

Methodologically, CondSEB-Net links sample-level adaptation, channel-level recalibration, and contour-level supervision within a controlled nnU-Net v2 baseline. CondConv addresses heterogeneous tumor morphology and texture, 3DSE attention reduces redundant channel responses under multi-scale feature compression, and boundary-aware supervision constrains blurred or irregular interfaces. The ablation results support this rationale, as each component improved the baseline and the combined model achieved the best overall performance among the ablation settings. Computational profiling further showed that these gains were obtained with limited additional overhead, mainly from CondConv kernel aggregation and 3DSE recalibration, while the boundary-aware loss did not increase inference-time parameters or FLOPs.

This study still has several limitations. First, the results were obtained from local five-fold cross-validation and were not submitted to the official online evaluation server. Second, although BraTS2020 and KiTS2019 are widely used public benchmarks, their standardized preprocessing cannot fully represent real-world clinical domain shifts, such as scanner variability, acquisition protocol heterogeneity, annotation differences, and unseen institutions. Therefore, the current results should be interpreted as benchmark-level evidence rather than direct evidence of clinical robustness, and external multi-center validation is still required. Third, the case-level failure analysis showed that isolated false-positive responses, tiny or spatially separated ET components, and residual boundary discrepancies in ambiguous low-contrast regions remain challenging. Future studies should further investigate lesion-size-aware learning, uncertainty estimation, and clinically guided post-processing to improve reliability in difficult cases.

Conclusions

This study proposed CondSEB-Net, a controlled enhancement of the nnU-Net v2 framework for 3D medical image segmentation. By incorporating CondConv-based dynamic convolution, multi-level 3DSE channel attention, and boundary-aware supervision, the proposed method improved adaptive feature representation and boundary alignment while preserving the self-configuring pipeline of nnU-Net v2. Experimental results on BraTS2020 and KiTS2019 showed that CondSEB-Net achieved superior or competitive Dice, HD95, and ASSD performance compared with baseline methods. In particular, fold-wise statistical analysis on BraTS2020 demonstrated statistically supported improvements in region overlap and boundary-related metrics. Ablation, qualitative comparison, and case-level failure analysis further confirmed the complementary effects of feature adaptation, channel recalibration, and boundary-aware learning, while also revealing remaining challenges in false-positive suppression, tiny enhancing lesion detection, and boundary ambiguity. Future work will focus on external multi-center validation under heterogeneous imaging protocols, cross-domain robustness evaluation, uncertainty calibration, and clinical reader assessment to further investigate the generalizability and clinical applicability of the proposed method.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the CLEAR reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0792/rc

Funding: This work was supported by the Doctoral Research Foundation of Guiyang University, China (No. GYUKY-2026).

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-0792/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Louis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, Hawkins C, Ng HK, Pfister SM, Reifenberger G, Soffietti R, von Deimling A, Ellison DW. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol 2021;23:1231-51. [Crossref] [PubMed]
2. Menze BH, Jakab A, Bauer S, Kalpathy-Cramer J, Farahani K, Kirby J, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Trans Med Imaging 2015;34:1993-2024. [Crossref] [PubMed]
3. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells W, Frangi A. editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Cham: Springer; 2015:234-41. doi: 10.1007/978-3-319-24574-4_28.
4. Cicek O, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: MICCAI; 2016:424-32.
5. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203-11. [Crossref] [PubMed]
6. Maier-Hein L, Eisenmann M, Reinke A, Onogur S, Stankovic M, Scholz P, et al. Why rankings of biomedical image analysis competitions should be interpreted with care. Nat Commun 2018;9:5217. [Crossref] [PubMed]
7. Kervadec H, Bouchtiba J, Desrosiers C, Granger E, Dolz J, Ben Ayed I. Boundary loss for highly unbalanced segmentation. Med Image Anal 2021;67:101851. [Crossref] [PubMed]
8. Hu J, Shen L, Sun G. Squeeze-and-excitation networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA. 2018:7132-41. doi: 10.1109/CVPR.2018.00745.
9. Yang B, Bender G, Le QV, Ngiam J. CondConv: conditionally parameterized convolutions for efficient inference. Advances in Neural Information Processing Systems 32 (NeurIPS 2019). Available online: https://proceedings.neurips.cc/paper/2019/hash/f2201f5191c4e92cc5af043eebfd0946-Abstract.html
10. Heller N, Isensee F, Maier-Hein KH, Hou X, Xie C, Li F, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 challenge. Med Image Anal 2021;67:101821. [Crossref] [PubMed]
11. Oktay O, Schlemper J, Le Folgoc L, Lee M, Heinrich M, Misawa K, Mori K, McDonagh S, Hammerla NY, Kainz B, Glocker B, Rueckert D. Attention U-Net: learning where to look for the pancreas. MIDL; 2018. Available online: https://openreview.net/forum?id=Skft7cijM
12. Chen J, Mei J, Li X, Lu Y, Yu Q, Wei Q, Luo X, Xie Y, Adeli E, Wang Y, Lungren MP, Zhang S, Xing L, Lu L, Yuille A, Zhou Y. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med Image Anal 2024;97:103280. [Crossref] [PubMed]
13. Hatamizadeh A, Tang Y, Nath V, Yang D, Myronenko A, Landman B, Roth HR, Xu D. UNETR: transformers for 3D medical image segmentation. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Waikoloa, HI, USA. 2022:574-84. doi: 10.1109/WACV51458.2022.00181.
14. Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B. Swin transformer: hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada. 2021:10012-22. doi: 10.1109/ICCV48922.2021.00986.
15. Hatamizadeh A, Nath V, Tang Y, Yang D, Roth HR, Xu D. Swin UNETR: Swin transformers for semantic segmentation of brain tumors in MRI images. In: Crimi A, Bakas S. editors. Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries. BrainLes 2021. Lecture Notes in Computer Science, vol 12962. Cham: Springer; 2022:272-84. doi: 10.1007/978-3-031-08999-2_22.
16. Woo S, Park J, Lee JY, Kweon IS. CBAM: convolutional block attention module. In: Ferrari V, Hebert M, Sminchisescu C, Weiss Y. editors. Computer Vision – ECCV 2018. ECCV 2018. Lecture Notes in Computer Science(), vol 11211. Cham: Springer; 2018. doi: 10.1007/978-3-030-01234-2_1.
17. Ma J, He Y, Li F, Han L, You C, Wang B. Segment anything in medical images. Nat Commun 2024;15:654. [Crossref] [PubMed]
18. Du Y, Bai F, Huang T, Zhao B. SegVol: universal and interactive volumetric medical image segmentation. In: Advances in Neural Information Processing Systems 37. Vancouver, Canada: NeurIPS; 2024. doi: 10.52202/079017-3516.
19. Wasserthal J, Breit HC, Meyer MT, Pradella M, Hinck D, Sauter AW, Heye T, Boll DT, Cyriac J, Yang S, Bach M, Segeroth M. TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images. Radiol Artif Intell 2023;5:e230024. [Crossref] [PubMed]
20. Shaker A, Maaz M, Rasheed H, Khan S, Yang MH, Shahbaz Khan F. UNETR++: Delving Into Efficient and Accurate 3D Medical Image Segmentation. IEEE Trans Med Imaging 2024;43:3377-90. [Crossref] [PubMed]
21. Pang Y, Liang J, Huang T, Chen H, Li Y, Li D, Huang L, Wang Q. Slim UNETR: Scale Hybrid Transformers to Efficient 3D Medical Image Segmentation Under Limited Computational Resources. IEEE Trans Med Imaging 2024;43:994-1005. [Crossref] [PubMed]
22. Liu J, Yang H, Zhou HY, Xi Y, Yu L, Yu Y, Liang Y, Shi G, Zhang S, Zheng H, Wang S. Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. Cham: Springer; 2024:615-25. doi: 10.1007/978-3-031-72114-4_59.
23. Isensee F, Wald T, Ulrich C, Baumgartner M, Roy S, Maier-Hein KH, Jaeger PF. nnU-Net Revisited: a call for rigorous validation in 3D medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. Cham: Springer; 2024. doi: 10.1007/978-3-031-72114-4_47.