The encoder consists of a hierarchical Swin Transformer backbone, which consists of extracting multi-scale contextual representations of the input MRI slices. Individual input MRI images are downsampled to 224 × 224 × 3 and broken down into 4 × 4 non-overlapping patches, which is the typical Swin Transformer tokenization approach. The result of this operation is: 2244×2244=56×56 resulting in 3136 visual tokens. The patches are, in turn, embedded into an embedding space, where a linear embedding layer is used to give a feature representation of the number of channels of an embedding dimension of 96.

The encoder has four layers of Swin Transformer that are made of window-based multi-head self-attention blocks and feed-forward layers. The shifted-window attention mechanism enables the network to model both local and global contextual dependencies while maintaining linear computational complexity. Between stages, patch merging operations are applied to reduce spatial resolution and increase feature dimensionality. This hierarchical design enables the encoder to progressively capture richer semantic representations of tumor structures. The spatial resolution and channel dimensions across encoder stages are summarized in Table 5.

This hierarchical feature extraction enables the model to acquire fine-tumor boundary information as well as global context relational information across the image.

Swin-DS-HAFUNetv2 uses an Enhanced Hierarchy Attention Fusion (E-HAFv2) module to minimize the semantic gap between the encoder and decoder representations. Unlike in the approaches of traditional skip connections, which simply concatenate encoder feature maps and decoder feature maps, E-HAFv2 uses attention-guided fusion to highlight the tumor-relevant areas and reduce the background noise.

As shown in Fig. 3, blocks of encoder and decoder characteristics with the same spatial resolution are initially optimized by Swin Transformer to enhance the strength of the contextual features. The maps of the features obtained are then concatenated, and a 1 × 1 convolution layer is applied to reduce the dimension. After this, a Squeeze-and-Excitation (SE) gating strategy is then used to recalibrate channel-wise feature responses adaptively, maximizing discriminative tumor features and reducing redundant/less relevant information.

Such a hierarchical approach of combining attention enhances the accuracy of tumor boundaries detection, especially in areas of the brain that are structurally complex, like the ventricles and cortical folds.

Swin-DS-HAFUNetv2 has a Contextual Bottleneck Transformer (CBT) in the deepest level of the network to obtain global contextual dependencies across the entire feature map. As illustrated in Fig. 4, CBT substitutes the conventional convolutional bottleneck with a transformer-based module, which incorporates:

Multi-Head Self-Attention (MHSA)

This design allows the model to capture long-range dependencies across tumor structures that may span large spatial regions in MRI images. Additionally, CBT enhances the stability of feature propagation and allows the network to separate diffuse tumor boundaries more effectively than would be suitably captured by local convolutional filters.

The decoder gradually recombines high-resolution segmentation maps with the help of hierarchical encoder attributes by using patch expansion and attention-fused features. The three major operations that are carried out by each decoder stage are:

Patch Expansion: Progressively, patch expansion operations (i.e., transposed convolution or interpolation) are carried out on the feature maps to restore the spatial resolution.

Deep supervision is used across various levels of the decoder to stabilise the training and enhance gradient propagation. The results of auxiliary segmentation (DS1, DS2, DS3) at intermediate resolutions are added to the total loss function, as shown in Fig. 5.

This multi-scale supervision strategy enhances the consistency of the segmentation and increases the speed of the network convergence.

Given the class imbalance that is typically seen in brain tumor segmentation problems, a hybrid loss function is used to trade region overlap and pixel-level classification performance.

The segmentation loss is defined as in Eq. (1):(1)ℒseg =ℒDice+ℒCE+ℒFocal Tversky 

To add deep supervision, the overall training goal can be a combination of the losses of results of the intermediate decoder as represented in Eq. (2):(2)ℒtotal=ℒDS1+0.6ℒDS2+0.3ℒDS3

This weighted representation will guarantee that the high-resolution predictions will dominate the optimization process, with intermediate outputs giving auxiliary gradient signals. The Adam optimizer that employs a ReduceLROnPlateau learning rate schedule is used to optimize the network. Table 6 provides a summary of the training setup of Swin-DS-HAFUNetv2.

The output segmentation of Swin-DS-HAFUNetv2 is considered to be the key element of the proposed explainability-aware framework, as it is the main input of the downstream classification phase. The framework provides classification models with pure tumor areas instead of unrelated background or imaging artifacts, as it explicitly isolates tumor areas and then classifies them. This tumor-centric representation is not only more accurate at making classification predictions, but also the visualizations provided by the corresponding Grad-CAM are more reliable. Specifically, Swin-DS-HAFUNetv2 generates a tumor mask that limits the spatial concentration of Grad-CAM explanations in choices, which ensures that saliency maps reflect anatomically significant parts of the tumor. The correct tumor segmentation is thus in direct correspondence with the explanation-consistency filtering strategy used in the ensemble classification phase, where attention maps are considered in anatomically relevant areas. This leads to saliency maps created that are more in line with tumor boundaries and have a lower rate of false activations, thus enhancing the interpretability and clinical accuracy of the prediction made by the model.

In general, Swin-DS-HAFUNetv2 offers a powerful and contextual segmentation backbone that can be seamlessly combined with an explainability-aware classification pipeline. The architecture shows a high level of scalability in terms of feature representation, interpretability, and performance of segmentation across the MRI slices in the BRISC2025 dataset, and this indicates the methodological and clinical viability of the proposed two-stage architecture.

The ECWMEv2 ensemble uses a pool of five transformer-based and convolutional neural architectures, which were chosen due to their variety of architectural biases and performance on medical imaging tasks. The model pool includes CoAtNet, ConvNeXt, Swin Transformer, MaxViT, and EVA02. CoAtNet is a combination of the convolution and self-attention features that help to compromise local spatial details and global dependencies. ConvNeXt is an updated design of CNNs that also considers design ideas of transformers but preserves the performance of convolutional operations. Swin Transformer applies window-based hierarchical self-attention, which allows learning fine-grained spatial structures in a scaled manner. MaxViT makes use of both block-based attention and grid-based attention to effectively combine local and global features. Lastly, EVA02 is a large-capacity vision transformer having trainable parameters with masked image models, which has high transfer learning in low-data medical imaging. All models are initialized with ImageNet-1K pretrained weights and fine-tuned on tumor-focused regions extracted from the BRISC2025 dataset using the output of the segmentation module. This guarantees both the localization of representation and the diagnostic relevance of the models. Table 7 gives a detailed description of the model specifications and training configurations.

Grad-CAM-based explanation consistency is used to assess whether candidate models have stable and clinically meaningful attention patterns, and only models that pass this test are included in the ensemble. Models that satisfy pre-determined consistency criteria are added to the final ensemble, and the other models are eliminated. The explanation consistency evaluation procedure, including similarity metrics, threshold criteria, and empirical justification, is detailed in Section 3.5.

Following the explanation, consistency filtering, the retained models are not treated equally. Instead, their contributions to the final prediction are weighted according to the degree of Grad-CAM stability observed during the filtering stage.

To further refine the decision-making process, ECWMEv2 uses a meta-classifier that is trained on the logits of the retained models, allowing a refined prediction using model-level feature combination. The logit vectors of each backbone are obtained as raw and concatenated together and provided as features to an XGBoost classifier. The meta-layer is a non-linear learner of relationships between predictions made by a model and can be used to resolve conflicts between base models adaptively. It is trained on a stratified validation split on a binary logistic loss and L2 regularization. Such a two-level fusion, being an explanation-filtered soft voting and then a meta-classification, allows the framework to achieve high generalizability without the loss of explainability.

The backbone models are optimized with the help of uniform hyperparameters and data augmentation plans. The output layers associated with the model are retrained in a multiclass classification and are trained by the cross-entropy loss. The training is done on NVIDIA A100 GPUs with PyTorch and mixed-precision so as to speed up convergence. XGBoost meta-classifier is implemented using the validation split with the maximum depth of 3, the learning rate of 0.05, and the L2 regularization weight of 1.0. Table 8 lists the training setup in each of the models in the ECWMEv2 pool.

To reduce the risk of overfitting and improve model generalization, several strategies were employed during training. First, transfer learning was applied by initializing backbone networks with ImageNet-1K pretrained weights. Second, it was augmented with vast amounts of data by using rotation, flipping, blur, and noise injection techniques. Third, early stopping based on patience of five epochs was used to avoid excessive training when the validation performance stopped improving. Further to guarantee the robustness and minimize sampling bias, five-fold cross-validation was used in experimental assessment, and the dataset was split into five folds, with the model being trained and evaluated on alternate splits. In the case of the segmentation network, deep supervision has been added to enhance gradient propagation and minimize overfitting at intermediate layers of features.

The explainability aspect of the suggested framework is achieved utilizing Grad-CAM as a post hoc interpretability approach, as opposed to utilizing internal transformer attention maps employed directly. Although transformer self-attention offers implicit interactions between features, attention weights are not always spatially interpretable as saliency regions of medical images. Thus, Grad-CAM is used to produce spatially localized heatmaps, which show areas of the tumor that are used in classification predictions.

In contrast to the traditional methods, when Grad-CAM is applied to visualize the model once the inference is made, the offered framework incorporates the concept of explainability within the model selection and the model ensemble building. In particular, Grad-CAM maps obtained using a candidate model are checked on controlled input perturbations. Models that give unstable or inconsistent explanations are not included in the ensemble. Therefore, explainability is a clear reliability metric used to filter and weight models, and interpretability is an inseparable part of the decision-making pipeline and not a purely descriptive post hoc visualization. The saliency-based attribution with Grad-CAM is employed to create spatially consistent attention maps, which are appropriate to interpret the classification of brain tumors. Grad-CAM, in contrast to SHAP or LIME, which feature-level attribution, offers anatomically aligned heatmaps, which are required in medical imaging tasks where regional continuity is important.

For given class c, Grad-CAM computes the importance weight of feature map k in the final convolutional layer as:(3)αkc=1Z∑i∑j∂yc∂Aijk

Here, yc is the score for class c, Ak is the kth feature map. Z is the spatial normalization factor. The Grad-CAM heatmap is then computed as:(4)Gc=ReLU(∑kαkcAk)

This produces a spatial saliency map indicating image regions contributing to the prediction. Each model in the classification ensemble produces Grad-CAM maps of clean and perturbed MRI inputs. Perturbations comprise: Rotation, Blur, and Noise injection simulating the clinically relevant variations of input. To measure the stability of the explanation, two measures are employed in the case of each pair of maps (clean Gc and perturbed Gp):

Intersection over Union (IoU) is a measure of spatial overlap: (5)IoU(Gc,Gp)=|Gc∩Gp||Gc∪Gp|≥0.6

The models that satisfy both thresholds (IoU≥0.6, Cosine Similarity ≥0.6) in 80% of the test samples are said to be explanation-consistent. The pass rate is calculated as in Eq. (7). This forms the Explanation Reliability Layer (ERL) in the ECWMEv2 framework, which filters out unstable models before ensemble construction. (7)Pass Rate=Samples with Stable CAMTotal Test Samples×100\% 

The retained models are not treated equally to ensure that there is fairness in the weighting of the models. Instead, each model is assigned a weight wm b proportional to its explanation consistency, calculated in Eq. (8):(8)wm=12(AvgIoUm∑jAvgIoUj+AvgCosm∑jAvgCosj)

The final class probability is obtained via explanation-weighted soft voting expressed in Eq. (9). Here, Z ensures normalization and pm(c∣x) is the softmax probability from model m for class c. (9)PECWMEv2(c∣x)=1Z∑m∈Msclectedwm⋅pm(c∣x)

Thresholds of IoU ≥ 0.6 and Cosine Similarity ≥ 0.6 with a pass rate ≥ 80% were determined empirically to be able to compromise between strictness and diversity in the ensemble. Table 9 demonstrates the impact of stricter and stricter thresholds: extremely conservative thresholds (0.8 or higher) will weed out all stable models, whereas moderately balanced thresholds will weed out none. It is not aimed at imposing the same attention maps but providing consistency that is both anatomically significant and clinically interpretable.

When all models were applied using the chosen thresholds, ConvNeXt, Swin Transformer, and EVA02 only had consistently stable and well-localized Grad-CAM heatmaps with tumor regions when perturbed. Conversely, CoAtNet as well as MaxViT were either diffuse or unstable and were eliminated in the final ensemble. The high consistency of the explanations by the average Grad-CAM IoU of 0.692 justifies the capability and transparency of the model. The quantitative analysis of the reliability of each model is provided in Table 10 to present the reliability of its explanation and retention.

The analysis has resulted in Grad-CAM being not just a visualization tool, but an important component of the ensemble in terms of filtering, weighting, and reliability of models. ECWMEv2 improves the readability, strength, and clinical credibility by ensuring that only models that are consistent and explanatory are used to make the predictions. The proposed explanation consistency filtering method is described in Algorithm 1.

Evaluation of the segmentation model (Swin-DS-HAFUNetv2) is measured in terms of existing spatial overlap measures. The Dice coefficient is used to measure the overlap between the predicted segmentation P and ground truth mask G, and it is given in Eq. (10):(10)Dice Coefficient =2|P∩G||P|+|G|

Eq. (11) is also used to compute the Intersection over Union or Jaccard Index:(11)IoU=|P∩G||P∪G|

To further analyze the model in terms of discriminative ability to divide into relevant areas, precision and recall are calculated using Eqs. (12) and (13):(12)Precision=TPTP+FP (13) Recall=TPTP+FN

Here,TP,FP, and FN denote true positives, false positives, and false negatives, respectively.

The classification module is evaluated by measuring its accuracy and classes. The total classification precision is in Eq. (14):(14)Accuracy=TP+TNTP+TN+FP+FN

To control the imbalance of classes, the F1-score is calculated as the harmonic mean of the precision and recall using Eq. (15):(15)F1−score=2×Precision ⋅ Recall  Precision + Recall 

To obtain a more in-depth understanding of the Swin-DS-HAFUNetv2 model, qualitative segmentation results on the validation set were viewed and clustered into good (≥0.90), borderline (0.75–0.85), and bad (≤0.70) results as illustrated in Fig. 7. This classification shows variability of segmentation performance and represents recurring patterns on clinical information.

Mean IoU (mIoU) per epoch was used to monitor the training process and was performed on a four-fold cross-validation of segmentation. The model demonstrated a steady convergence of the model along the folds, as demonstrated in Fig. 8, and optimality was reached at epoch 25; the output thereafter stabilized. The mean mIoU of the architecture was 0.865 ± 0.16, which validates the consistency of the architecture on varying validation splits as well as the efficiency of deep supervision and integration of attention.

Table 11 demonstrates the results of segmentation by cross-validation folds. The outcomes show a consistent performance, and the Dice scores are between 0.9780 and 0.9785. The model is relatively stable and reliably performed across all folds with an average precision of 0.9363 and a recall value of 0.9040 to 0.9091.

Further statistical reliability testing was done on the proposed model, Swin-DS-HAFUNetv2, through cross-validation analysis of the confidence interval. The mean of the Dice score was 0.9782, and the standard deviation was 0.0003; hence, the 95% confidence interval of the score was 0.9779–0.9785. The small range suggests that the performance of segmentation is close to the same value in the various validation splits, which proves the performance gains are statistically constant and not simply by chance.

To quantify the individual contributions of the E-HAF and the E-CBE modules, a systematic ablation study was conducted, as summarized in Table 12. The baseline Swin DS HAFUNetv2 model without either component achieved a weighted mIoU of 85.2%. The usage of E-HAF in isolation enhanced multi-scale feature integration, achieving consistent improvement in all tumor types and weighted mIoU to 86.5%. The added E-CBE further contributed to better contextual modeling of the bottleneck at the global scale, the weighted mIoU increased to 87.1%, and induced significant positive changes for glioma and meningioma cases. The overall performance of the model was the greatest when both E-HAF and E-CBE were used, and the glioma group had the highest increase. These findings indicate that E-HAF and E-CBE enhance each other in terms of benefits, and when they are combined, they provide a significant advantage to tumor boundary demarcation and contextual insight, especially of more complex and heterogeneous tumor structures. The results confirm the idea that E-HAF enhances the feature fusion process of different scales, and E-CBE enhances the bottleneck presentation. Together, they produce significant mIoU improvements, especially on the more complicated glioma subtype.

The proposed Swin-DS-HAFUNetv2 model was tested and compared to some of the current segmentation models. Recent transformer-based segmentation models such as TransUNet, Swin-UNet, and UNETR have demonstrated strong performance in medical image segmentation tasks [66]. However, many of these architectures are primarily designed for volumetric multimodal MRI datasets such as BraTS. Since the BRISC2025 dataset contains single-slice T1-weighted images, direct comparison with these models requires substantial architectural adaptation. As shown in Table 13, the proposed method significantly outperforms existing studies, particularly in glioma, meningioma, and pituitary cases.

Each backbone model was trained on the classification task. The results in Table 14 indicate that Swin Transformer and ConvNeXt outperformed the rest in terms of both predictive and explanation consistency.

Explanation consistency was assessed under noise, rotation, and contrast jitter. Swin, ConvNeXt, and EVA02 passed the minimum threshold (IoU ≥ 0.6, cosine similarity ≥ 0.6, pass rate ≥ 0.8), as shown in Tables 10 and 15. Qualitative overlays of Grad-CAM on tumor regions of a sample data in Fig. 9 demonstrate moderate overlap (Dice = 0.587, IoU = 0.416), confirming alignment between model attention and pathology.

Fig. 10 contains Grad-CAM visualization with varying perturbations of inputs, which clearly shows that some models are more stable to explanations. It is also evident that Swin Transformer and ConvNeXt always have well-localized and strong attention maps, which do not change in brightness, noise, and rotation perturbation, meaning that they can consistently focus on tumor localities. EVA02 will show relatively wider and noisier activations, but they still are semantically meaningful and would yield valuable global contextual data. CoAtNet, in contrast, exhibits fragmented and dispersed attention, which is highly sensitive to perturbations, whereas MaxViT generates weak or diffuse activations that are of low anatomical relevance. These findings show that Swin Transformer, ConvNeXt, and EVA02 are the only ones that meet the quantitative explanation-consistency criteria (IoU ≥ 0.60, cosine similarity ≥ 0.60, and pass rate ≥ 0.80) and are thus included in the final ensemble.

The ECWMEv2 framework proposed was compared to the explanation-weighted soft voting and single classifiers. As shown in Table 16, ECWMEv2 employing an XGBoost meta-learner achieved superior class-balanced performance, attaining a macro-F1 score of 0.9867 with an accuracy of 0.9917. In comparison, explanation-weighted soft voting achieved lower accuracy (0.9863) and lower macro-F1 (0.9734), indicating reduced robustness in handling inter-class variability.

To verify that the performance improvement achieved by ECWMEv2 is statistically meaningful, statistical significance testing was conducted using five-fold cross-validation results. The proposed ECWMEv2 ensemble achieved an average accuracy of 0.9917 ± 0.0046, compared with 0.9863 ± 0.0031 for explanation-weighted soft voting. Similarly, the Macro-F1 score improved from 0.9734 ± 0.0062 to 0.9885 ± 0.0075. The stability of ECWMEv2 is also attested by the five-fold results of the cross-validation in Table 17, with accuracy ranging between 0.9897 and 1.0000, and F1 scores ranging between 0.9789 and 1.0000. The mean of folds (accuracy = 0.9917, macro-F1 = 0.9867) indicates a high generalization and a low variance, which proves that the model selection based on explanation consistency is part of the consistent and high accuracy of classification.

Table 18 summarizes the effect of each component in the ECWMEv2 classification framework. Using the best-performing single model (Swin Transformer) as a baseline, we observed an accuracy of 98.71% and a macro-F1 of 98.12%, with high Grad-CAM localization consistency (Mean IoU = 0.722). When the Explanation Reliability Layer (ERL) was removed, and all models were included without Grad-CAM filtering, performance dropped to 98.42% accuracy and 97.81% macro-F1, with a significantly lower explanation quality (Mean IoU = 0.461) due to the inclusion of unstable models like CoAtNet and MaxViT.

The fusion of explanation-based and uniform voting to the filtered models resulted in 98.83% accuracy and 98.29% macro-F1, indicating that explanation-weighted fusion enhances calibration and class balance. The overall performance is the best with ECWMEv2 architecture, both Grad-CAM consistency filtering (ERL) and explanation-weighted soft voting, with the final accuracy of 99.17, the macro-F1 of 98.67, and the Mean IoU of 0.692. These results confirm that the model selection that reduces problems with an explanation and focuses on the ensemble fusion helps to increase the reliability of prediction and the interpretability of the model in brain tumor classification.

Besides predictive performance, the computational complexity of the proposed framework was evaluated based on model parameters, training time, and inference latency.