Multi-Modal Brain Tumor Segmentation with ResSE-LKA Enhanced U-Net Architecture

Brain tumors, particularly gliomas, remain among the most life-threatening neurological diseases due to their highly infiltrative nature, pronounced structural heterogeneity, and high recurrence rates. In parallel, the global incidence of cancer has continued to rise in recent years, posing significant challenges to healthcare systems worldwide. Accurate identification and delineation of brain tumor subregions from magnetic resonance imaging (MRI) play a pivotal role in numerous clinical applications, including surgical planning, image-guided therapeutic interventions, longitudinal tumor monitoring, and radiotherapy planning [1]. Moreover, precise tumor subregion segmentation is increasingly recognized as a fundamental prerequisite for quantitative imaging analysis, which can further contribute to prognostic assessment and overall survival prediction in patients [2].

Multimodal MRI is widely employed for brain tumor diagnosis and evaluation due to its ability to provide rich anatomical and tissue-specific information, with each imaging modality emphasizing distinct tumor characteristics such as enhancing tumor regions, necrotic cores, and peritumoral edema [3]. However, manual delineation of tumor subregions on multimodal MRI is a time-consuming and labor-intensive process that requires substantial expertise and is highly dependent on the individual experience of radiologists. In routine clinical practice, this task is predominantly performed through qualitative visual assessment, which often leads to inter-observer variability and limits its feasibility in large-scale clinical settings [4]. These limitations highlight the urgent need for automated, accurate, and reliable brain tumor segmentation (BraTS) methods that can assist clinicians, reduce workload, and improve consistency in medical image analysis. This demand for automation mirrors similar trends observed across other diagnostic domains, where deep learning-based frameworks have demonstrated strong potential in automated fault detection and predictive diagnosis from sensor signal data [5].

Early brain lesion segmentation methods mainly relied on traditional image processing techniques such as thresholding, edge detection, and clustering. For example, Sulaiman et al. [6] investigated the use of clustering-based algorithms for segmenting brain MRI images, proposing a method where an adaptive fuzzy K-means (AFKM) algorithm is used to partition the MRI intensity space and group similar tissue types, enabling automatic separation of tumor regions from healthy brain tissues. However, the performance of clustering-based methods can be sensitive to noise, the choice of initial cluster centers, and image intensity inhomogeneity, which affect the accuracy and consistency of segmentation results. While these methods could achieve reasonable performance under controlled conditions, they depended heavily on hand-crafted features and manually selected parameters, leading to a noticeable degradation in accuracy when tumor shapes became complex or MRI images were affected by noise.

With the emergence of deep learning, segmentation models began to learn feature representations automatically in an end-to-end manner. Fully Convolutional Networks (FCNs) introduced a pixel-wise segmentation framework by eliminating fully connected layers [7]; however, repeated downsampling operations often caused the loss of spatial details, resulting in blurred boundaries and limited ability to detect small tumor regions. Convolutional neural networks have since become central to medical image segmentation, among which the U-Net [8] is one of the most widely adopted architectures. Its encoder–decoder structure with skip connections effectively combines semantic context with fine-grained spatial information, making it suitable for medical tasks with limited annotated data. Building on this framework, numerous U-Net-based variants have been proposed and evaluated on the BraTS benchmark [9, 10], which has become a standard dataset for the comparative evaluation of BraTS methods.

Despite these advances, BraTS from MRI remains difficult due to several inherent factors. First, tumor subregions show large variations in appearance across different MRI modalities and across patients, leading to highly heterogeneous intensity and texture patterns. Second, tumor boundaries are often irregular and unclear, particularly in edematous regions where the transition between tumor tissue and healthy brain is gradual rather than sharp. Third, the strong imbalance between tumor and background voxels can cause models to favor the dominant background class, reducing sensitivity to small tumor regions. Finally, variations in scanners and imaging protocols across institutions introduce intensity inconsistencies that can significantly affect model robustness and generalization.

To address these challenges, recent research has focused on enhancing the representational capacity of segmentation networks. A key direction involves strengthening feature representation and recalibration. The Squeeze-and-Excitation (SE) block [11] introduced a channel-wise attention mechanism that adaptively recalibrates feature responses, showing promise in medical tasks. ResUNet incorporates residual learning into the U-Net architecture, enabling deeper feature extraction while maintaining stable gradient propagation. The use of residual blocks allows the network to effectively aggregate multi-level contextual information and enhance semantic representation, thereby improving its ability to model complex tumor structures and capture both local details and broader contextual cues [12]. The recent Large Kernel Attention (LKA) module offers an efficient decomposition of large convolutional kernels into depthwise and pointwise components, effectively modeling global context with manageable computational cost [13]. Furthermore, explicitly modeling multi-scale contextual information at the network bottleneck, where features are most abstract, can enhance semantic understanding before upsampling [14].

In this paper, we propose a novel network architecture, ResSE-LKA-UNet, designed to tackle the aforementioned limitations of existing BraTS methods. Our contributions are threefold:

• Architectural Design: We propose an enhanced 2D U-Net integrating ResSE blocks for channel-aware feature extraction, LKA modules for modeling long-range spatial dependencies, and a Context Enhancement (CE) module to fuse multi-scale contextual information.
• Optimization Strategy: A hybrid weighted Cross-Entropy and adaptive Tversky Dice loss with deep supervision is employed to address class imbalance and improve training stability.
• Evaluation: Experiments on a subset of the BraTS 2021 dataset demonstrate competitive performance across major tumor subregions (ET, ED, NCR), with qualitative results showing improved delineation of edematous and complex tumor structures compared to the baseline U-Net.

3.1 Datasets

Experiments were conducted on the BraTS 2021 (2040 cases) dataset across four MRI modalities (T1, T1ce, T2, and FLAIR) [3], as illustrated in Figure 6. All experiments follow a limited-data setting. Specifically, 10% of the available cases were randomly selected using a fixed random seed of 42. These cases were first divided into training and validation sets without overlap to avoid data leakage. The selected 204 cases were split into 163 cases for training and 41 cases for validation. Preprocessing, 2D axial slice extraction, and data augmentation were then performed on the resulting splits. After training, the model was further applied to several unseen cases from the remaining data for qualitative prediction analysis. We then adopted a 5-fold cross-validation protocol.

6.png

Figure 6. Multimodal magnetic resonance imaging (MRI) visualization and tumor segmentation from (BraTS) 2021. Modalities: FLAIR, T1, T1ce, T2, ground-truth mask, and overlay. Tumor sub-regions: NCR, edema, enhancing tumor

3.2 Preprocessing

The BraTS input data are 3D multimodal MRI volumes comprising FLAIR, T1, T1ce, and T2 sequences. To reduce computational cost and memory usage, a 2D slice strategy is adopted, where axial slices are extracted from each 3D volume for segmentation. For each subject, every modality is processed independently. The images are first cropped to remove redundant background, then resized to 128 × 128. Intensity normalization is subsequently performed on the non-zero voxels of each modality volume: intensities are clipped to the [0.5^th, 99.5^th] percentile range and then standardized using z-score normalization. The segmentation masks are not modified, as shown in Figure 7.

7.png

Figure 7. Visualization of the preprocessing steps

For experiments using contrast enhancement, CLAHE is additionally applied to all image modalities in a slice-wise manner with a clip limit of 2.0 and a tile grid size of 8 × 8. This step is intended to improve local contrast and facilitate the delineation of subtle tumor regions and unclear boundaries.

To improve generalization under the limited-data setting, the following online augmentation transforms are applied to each training slice during loading. The augmentation transforms and corresponding parameters are summarized in Table 1.

Table 1. Data augmentation parameters used in model training

3.3 Implementation details

All experiments were conducted on a Windows 11 system equipped with an Intel Core i7-13620H processor and an NVIDIA RTX 4060 GPU with 8 GB VRAM. The proposed model was developed based on the nnUNet baseline [18], and all comparative experiments were performed under matched settings to ensure a fair comparison.

Table 2. Hardware and software configuration used in the experiments

The model was implemented in Python using the PyTorch framework and trained for 50 epochs with a batch size of 8. CUDA 12.7 and cuDNN were used to accelerate GPU computation (as detailed in Table 2).

3.4 Loss function

The proposed model is optimized using a hybrid loss function that combines weighted CE loss and an adaptive Tversky-based Dice loss, further enhanced by deep supervision at multiple decoder stages. This design aims to jointly address pixel-wise classification accuracy, region-level overlap, and training stability under severe class imbalance.

$\mathcal{L}_{\text {total }}=\sum_{i=0}^K \lambda_i \mathcal{L}_i$,                   (9)

where, $K=3$ denotes the number of auxiliary deep supervision outputs, $\lambda_i$ represents the weighting factor of the $i$-th output, and $\mathcal{L}_i$ is the loss computed at that output.

The deep supervision weights are empirically set as:

$\left[\lambda_0, \lambda_1, \lambda_2, \lambda_3\right]=[1.0,0.4,0.3,0.2]$                     (10)

1. Weighted Cross-Entropy Loss

To perform pixel-wise multi-class classification, the weighted Cross-Entropy loss is defined as [8]:

$\mathcal{L}_{C E}=-\sum_{c=0}^{C-1} w_c y_c \log \left(\hat{p}_c\right)$                  (11)

where, $C=4$ corresponds to the segmentation classes: background, ET, ED, and NCR. Here, $y_c$ denotes the one-hot encoded ground truth, and $\hat{p}_c=\operatorname{softmax}(z)_c$ represents the predicted probability for class $c$.

To mitigate the impact of class imbalance inherent in BraTS, class weights are assigned as:

$w=[0.1,1.0,1.0,1.0]$                   (12)

The reduced weight for the background class prevents it from dominating the optimization process and encourages the network to focus on tumor-related regions.

2. Adaptive Tversky Dice Loss

While Cross-Entropy loss ensures discriminative pixel-wise learning, it does not explicitly optimize region overlap. Therefore, we incorporate a Dice loss based on an adaptive Tversky formulation, defined as [19]:

$\mathcal{L}_{\mathrm{T}}=1-\frac{1}{C} \sum_{c=0}^{C-1} \frac{T P_c}{T P_c+\alpha_c F P_c+\beta_c F N_c+\epsilon}$                    (13)

where, $T P_c, F P_c$, and $F N_c$ denote the true positive, false positive, and false negative counts for class $c$, respectively, and $\epsilon=10^{-5}$ is a smoothing constant.

The weighting coefficients $\alpha_c$ and $\beta_c$ are adaptively determined according to the prediction error distribution:

$\alpha_c=\operatorname{clamp}\left(\frac{F P_c}{F P_c+F N_c}, 0.2,0.8\right), \beta_c=1-\alpha_c$                (14)

This adaptive mechanism dynamically balances precision and recall, which is particularly important in medical image segmentation, where false negatives are often more clinically critical than false positives.

3. Final Loss

To improve gradient propagation and accelerate convergence, deep supervision (ds) is applied at multiple decoder levels [20]. The total loss is thus expressed as:

$\mathcal{L}_{\text {total }}=\sum_{i=0}^3 \lambda_i\left(\mathcal{L}_{C E}^{(i)}+\mathcal{L}_{\mathrm{T}}^{(i)}\right)$                    (15)

To be more specific:

$\mathcal{L}_{\text {total }}=1.0 \mathcal{L}_{\text {main }}+0.4 \mathcal{L}_{\mathrm{ds} 4}+0.3 \mathcal{L}_{\mathrm{ds} 3}+0.2 \mathcal{L}_{\mathrm{ds} 2}$                      (16)

All auxiliary outputs are upsampled to the original resolution before loss computation. The decreasing weights reflect the progressively coarser semantic information at deeper decoder stages.

3.5 Evaluation metrics

The Dice coefficient, a statistic used to assess the similarity between the segmentation results produced by our algorithm and the ground truth labels, is calculated as:

Dice $=\frac{2 \cdot \mathrm{TP}}{\mathrm{FP}+2 \cdot \mathrm{TP}+\mathrm{FN}}$                  (17)

where, TP denotes true positives, FP false positives, and FN false negatives.

The 95th percentile Hausdorff distance (HD95) quantifies the spatial discrepancy between the segmented tumor and the ground truth. It is defined as:

HD95 $=H_D(X, Y)=\max \left(d_{x y}, d_{y x}\right)$                      (18)

where, $d_{x y}$ and $d_{y x}$ represent the 95th percentile of the distances from points in set X to their nearest neighbors in set Y , and vice versa. A lower HD95 value signifies a closer spatial agreement between the predicted and actual tumor boundaries.

In this study, we proposed a novel segmentation framework for brain tumor subregion delineation on the BraTS 2021 dataset. The proposed model was designed to improve the representation of tumor structures by combining multi-scale feature extraction and enhanced spatial learning. Experimental results demonstrate that our approach achieves more stable and accurate performance compared to baseline methods such as U-Net and Attention U-Net.

In particular, the proposed model shows clear improvements in challenging tumor regions, especially edema (ED), which often presents blurred boundaries and diffuse patterns. Moreover, the model is able to better detect small lesions, where traditional architectures frequently fail to produce valid predictions. Both quantitative results and qualitative comparisons confirm the reliability and robustness of our method across different tumor sizes and complex cases.

Although the proposed method achieves promising results, several directions can be explored in future research. First, we plan to evaluate the model on larger datasets and additional multi-center MRI benchmarks to further validate its generalization ability. Second, incorporating transformer-based attention mechanisms or more advanced feature fusion strategies may further enhance segmentation accuracy, especially in extremely small or heterogeneous tumor regions.

In addition, future work could investigate semi-supervised or weakly supervised learning approaches to reduce the dependence on fully annotated data, which is often limited in medical imaging. Finally, optimizing the computational efficiency of the model will also be an important step toward real-time clinical deployment.