WorksComputer Science
Quality-Governed Deep Learning Network for Brain Tumor MRI Segmentation
Abstract
The segmentation of brain tumors from magnetic resonance imaging (MRI) is an essential step for computer-aided neuro-oncology, treatment planning and follow-up. While effective encoder-decoder representations are achieved by U-Net and Attention U-Net, the skip sections of the network can generate weak-contrast or noisy responses in the case of heterogeneous boundaries of the tumor. In this paper, we introduce a Quality-Governed Deep Learning Network (QGDL-Net) that can be used for binary tumor segmentation from FLAIR MRI slices. The proposed model is able to predict the slice quality based on contrast and SNR statistics and to fine-tune the encoder features prior to decoding by learnable quality gates. We used the 80:20 train-validation split for experiments with BraTS 2021. QGDL-Net achieved a Dice of 0.9747, IoU of 0.9506, precision of 0.9755, recall of 0.9739, F1 of 0.9747 and accuracy of 0.9982, outperforming U-Net and Attention U-Net. The EigenCAM maps and segmentation overlays provide better tumor localization and interpretable activation around the lesion regions.
Accepted at the 7th International Conference on Computational Vision and Bio Inspired Computing (ICCVBIC-2026). © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Introduction
Accurate segmentation of glioma and other brain tumor regions from MRI supports tumor volume estimation, radiation planning, surgical decision support and longitudinal response assessment. In neuro-oncology, MRI is preferred due to the complementary tissue contrast, pattern of edema and enhancing tumors with FLAIR, T1, T1ce and T2 sequences. The BraTS project unified benchmark data and assessment methods for this problem, allowing the algorithms to be compared in a repeatable way [1, 2, 3]. But, the morphology of the tumors is challenging due to different sizes, locations, intensities, necrotic content and sharpness of the tumor margins.
Brain tumor segmentation has been revolutionised by deep learning, from hand-crafted pipelines to trainable convolutional and transformer-based systems. It has been shown that encoder-decoder models with skip connections are good biomedical segmentation baselines [4, 5, 6] with the aid of U-Net, 3D U-Net and nnU-Net. Subsequent works adopted attention gates, cascaded networks, multiscale feature fusion and transformer modules, which further enhanced the boundary localization and global context modelling [7, 8, 9, 10, 11]. All these advances, however, fail to account for the specific nature of the response – whether it is high-contrast tumor evidence or low-quality, background-dominated responses – in the current slice.
This work presents a quality-controlled and quality-governed segmentation network called QGDL-Net, which evaluates the quality of the slices and adjusts the features of the encoder according to their quality, before passing them on to the decoder for fusion. According to the recent literature, feature quality, receptive field design, attention and explainability are crucial for a reliable medical segmentation [12, 13, 14, 15, 16]. The main contributions are: (i) a quality descriptor based on contrast and signal-to-noise statistics; (ii) channel-wise quality gates, placed inside skip pathways as in U-Net; (iii) comparative tests against U-Net and Attention U-Net; and (iv) visual validation by segmenting and overlaying on the basis of the metrics and plotting a metric graph, using EigenCAM. Figure 1 shows a representative BraTS2021 FLAIR MRI slice with its corresponding segmentation mask and superimposed tumor region.

Figure 1. Dataset view showing FLAIR input, tumor segmentation and superposition.
The motivation for the proposed method is twofold. First, MRI slices do not equally contribute to the learning due to the presence of some slices having large and distinct tumor regions, while others have weak borders, partial appearance or the dominance of tissue in the background. Second, the standard skip connections take low-level spatial information directly to the decoder, and that can be helpful for retaining the structure, but noisy activations can also be carried over. A quality-governed gate is a simple and trainable gate that regulates this transfer.
The methodological comparison is done keeping all the pre-processing steps, optimizer, and validation split the same for all the networks. The evaluation employs both region-overlap metrics and visual evidence as segmentation models are evaluated not just on a numerical basis, but also on the basis of the anatomical plausibility of the mask predicted. The goal of practical computer-aided diagnosis is to have an accurate, stable and explainable segmentation model.
The present work fulfills these requirements in three complementary parts: (1) Segmentation backbone with added performance, (2) Internal quality module and (3) EigenCAM visualization. The analysis shows that QGDL-Net can boost the segmentation quality by selecting more informative encoder features for reconstruction, confirming the claim.
The rest of this paper is divided into the following way. § 2 discusses some of the related literature. § 3 proposes the system model and elaborates on the offered QGDL-Net, and § 4 presents the analysis and results of the experiments. Lastly, § 5 is a conclusion of the paper, including the future research directions.
Literature Survey
The segmentation benchmark and the stress on standardized evaluation in terms of Dice and Hausdorff measures [1] were highlighted. Bakas et al. also annotated the glioma images and assisted the evolution of BraTS algorithms over the years [2]. Baid et al. enhanced the benchmark in BraTS 2021 using a big multi-institutional cohort and a radiogenomic classification task [3]. These datasets could be used to perform consistent algorithmic comparisons and show the presence of class imbalance and heterogeneous tumor boundaries.
U-Net is a symmetric encoder-decoder network with skip connections that was proposed in [4] for biomedical segmentation. Çiçek et al. furthered this idea by creating volumetric segmentation with 3D U-Net [5]. nnU-Net, proposed by Isensee et al., automatically made many design decisions and served as a good reference method for biomedical segmentation [6]. U-Net introduced attention gates to bypass irrelevant areas, enabling the decoder to focus on relevant structures [7].
The current techniques are not limited to conventional convolutional encoders. Transformer blocks were added to brain tumor segmentation by TransBTS, exploiting global dependencies [8]. UNETR proposed tokens for image patches for volumetric segmentation using a transformer [9], and SwinBTS and nnFormer investigated hierarchical attention and the design of a volumetric transformer [10, 11]. Later published papers in the journal introduced effective variants of U-Net enhanced by transformers, interpretable transformer hybrids, multiscale attention U-Net, LIU-Net and meta-transformer nnU-Net for brain tumor MRI [12, 13, 14, 15, 16]. While these works inspire the proposed quality-governed design, QGDL-Net differs in that it explicitly includes a quality score of the slices to modulate channels of the encoder prior to skip fusion.
The first set of related works is that of convolutional encoder-decoder segmentation. In the case of small medical structures, U-Net has developed the concept of concatenating high-resolution features of the encoder with features of the decoder, which is useful in this case. 3D U-Net expanded the concept to volumetric learning, which is not only more memory-intensive but also demands careful sampling of patches [5]. Systematic configuration of normalization, spacing, patch size, loss and augmentation showed to be sometimes as effective as picking a fixed architecture by hand [6] in the case of nnU-Net.
The second group presents attention and transformer modules. U-Net suppresses irrelevant skip features by means of attention using gating coefficients [7]. Self-attention is capable of learning global contexts which are hard for local convolutions [8, 9] by themselves, as is done by TransBTS and UNETR. To alleviate the computation load of attention, SwinBTS and nnFormer employ hierarchical windows or volumetric attention structure [10, 11]. These methods enhance representation learning, but they typically do not obtain an explicit image quality score prior to skip fusion.
The third class is models published recently in the journal that focus on efficient, interpretable and transferable segmentation. ETUNet is based on the transformer-enhanced U-Net design for 3D tumor segmentation [12]. Explainable hybrid transformer frameworks report that interpretability can enhance self-confidence in segmentation choices [13]. Recently, feature fusion, lightweight design and pretrained segmentation knowledge are key areas of ongoing research [14, 15, 16], with multi-scale attention U-Net, LIU-Net and meta-transfer learning strategies highlighting these. In this area, QGDL-Net is seen as a quality-controlled feature-fusion approach.
Based on the literature reviewed, three gaps are found. First, networks with high capacity can have high Dice scores, but can be very expensive to compute. Second, attention mechanisms tend to learn about the importance of the spatial location but may not explicitly learn the quality of the current slice at the global level. Third, some methods claim numerical improvements, but do not provide a concise mathematical description of the impact of the suggested feature regulation on noise. The proposed QGDL-Net overcomes these limitations with a lightweight quality estimator, channel-wise gates and a bounded-noise argument.
Proposed Methodology
Dataset and Experimental Design
The BraTS 2021 Task 1 training data was used in the study. Volumes and masks corresponding to the different FLAIR were loaded from NIfTI files, as were masks corresponding to the expert’s segmentation. The volume segmented for each patient was transformed to a binary mask by giving all labels of the tumor to one class. The volumes were z-score normalized, those slices of the axial slices with a tumor ratio > a minimum threshold were selected, and the images were resized to pixels. The created slices were divided into 80% training and 20% validation sets. The FLAIR sequence was implemented in the prototype due to its ability to clearly show hyperintense regions of many tumors and edema. While multimodal MRI can help improve subregion segmentation, a single-modality setting is valuable to test whether the proposed quality gate can aid in enhancing feature transfer within a constrained input setting. The dataset characteristics and preprocessing details used in this study are summarized in Table 1.
| Item | Setting |
| Dataset | BraTS 2021 Task 1 |
| Input | FLAIR MRI |
| Mask | seg > 0 |
| Size | |
| Patients | 50 |
| Split | 80:20 |
| Batch | 16 |
| Platform | Kaggle GPU |
Table 1. Dataset and preprocessing summary.
All compared networks underwent a similar pre-processing procedure. Normalization of each volume was performed by dividing by the mean and standard deviations of the respective volumes. Slices lacking adequate area of tumour were discarded to prevent training of too many almost empty masks. The same image-mask pairs are selected for U-Net, Attention U-Net and QGDL-Net to avoid the impact of data imbalance and different sampling, so that the differences in performance are due to the differences in structure. The task was defined as a binary whole-tumor segmentation. BraTS labels contain many tumor subregions, but in this study, all the labels of the tumor were summarized to a single foreground class. For this purpose, this decision is suitable to show the core quality-governed feature mechanism and isn’t mixed with multiclass loss weighing. The same gate could be expanded to improve tumor segmentation, tumor core and whole tumor segmentation in the future. Table 2 presents the architectural configuration of the compared segmentation models.
| Model | Core mechanism | Role |
| U-Net | Skip encoder-decoder | Baseline |
| Attention U-Net | Attention-gated skips | Attention baseline |
| QGDL-Net | Quality-gated skips | Proposed |
Table 2. Compared models.
Problem Formulation
Let be a normalized FLAIR slice, and be the tumor mask. The objective is to learn , where is a pixel-wise tumor probability map. The final mask is obtained by applying a threshold of 0.5 to .
The encoder progressively maps the input image into deeper feature spaces, where spatial resolution decreases and semantic abstraction increases. The decoder restores spatial resolution using transposed convolutions and feature concatenation. This design allows coarse contextual information from deep layers to be combined with fine anatomical details from early layers. The limitation is that skip concatenation is unconditional, which motivates quality-governed regulation.
Quality-Governed Feature Recalibration
QGDL-Net computes a scalar quality descriptor for each slice using contrast and signal-to-noise information:
For encoder feature with channels, a learnable gate maps into channel weights:
Here is the sigmoid function, is ReLU and denotes channel-wise multiplication. The decoder receives the quality-recalibrated features , and . Since , the gate cannot amplify noisy encoder energy. If , where denotes lesion signal and denotes disturbance, then . Thus, quality gating bounds non-informative feature energy while the loss encourages channels useful for tumor prediction to keep larger weights.
During the training phase, the parameters of the gates are adjusted by gradients of the Dice and BCE losses to better preserve channels that are associated with the tumor overlap, while the influence of channels that are repeatedly assigned false positives is reduced. This mechanism is easier to compute as only a scalar quality input and two fully connected layers are used at each level of the gate. Quality Gates are different than attention gates. Attention U-Net emphasizes the spatial relevance of the features in the decoder and features in the encoder, and QGDL-Net recalibrates the channels based on the overall quality of the slice. Thus, the quality mechanism can be viewed as reliability control of features, while the spatial attention mechanism can be viewed as a selection of locations. To maintain the compact architecture, the proposed model is based on the former. The overall architecture of the proposed QGDL-Net is illustrated in Figure 2.

Figure 2. Proposed QGDL-Net architecture with quality estimator and quality-gated skip connections.
Loss Function and Evaluation Metrics
All models were optimized using a compound loss:
where denotes the binary cross-entropy (BCE) loss and denotes the Dice loss defined in (8).
Performance was evaluated using Dice, IoU, precision, recall, F1-score, accuracy, ROC analysis and confusion matrices. EigenCAM was used to visualize activation regions of the proposed network. The training configuration and experimental settings are listed in Table 3.
| Parameter | Value |
| Optimizer | Adam |
| Loss | BCE + Dice |
| LR | 0.001 |
| Threshold | 0.5 |
| Epochs | 10/25/50 |
| Criterion | Best Dice |
Table 3. Training configuration.
Implementation and Reproducibility Details
The implementation was created with PyTorch and run in a Kaggle GPUs environment. To ensure repeatability, the same dataset object, data split, mini-batch Algorithm 1 Quality-Governed Training and Inference
Require: MRI slices , masks , models , epochs
Ensure: Trained segmentation model and probability mask
Normalize FLAIR volume and extract tumor-containing axial slices.
for epoch to do
for each mini-batch do
Compute , and quality score .
Extract encoder maps , , .
Estimate using quality gates.
Recalibrate and decode .
Update by minimizing .
end for
end for
return binary mask during inference.
size and loss function were used in three models. According to the validation Dice, validation model checkpoints were saved separately. To let the gate and attention parameters remain stable, the Attention U-Net and QGDL-Net were trained for longer epochs while the baseline U-Net was evaluated for fewer epochs, as it converged rapidly. The chosen evaluation protocol was carefully chosen to be simple and clear. Each validation mask was segmented into foreground and background pixels to calculate precision, recall, F1-score, accuracy and IoU. The Dice measure was calculated using the intersection of the predicted binary mask and the ground truth mask. Before thresholding, the raw probability maps used to compute ROC curves were used to show how well the pixels from the tumor and background could be separated across various thresholds.
While the model is learnt in slices of 2D images, it retains the essence of learning features guided by quality. The same quality gate can be generalized to 3D feature tensors by using 3D convolutions and calculating from a volume/patch. In this work, the adopted score for quality is deliberately light. It integrates the contrast with signal-to-noise properties, instead of using an extra neural quality evaluation network. This will not result in the proposed method becoming significantly bigger than the baselines. Furthermore, it simplifies the mechanism to interpret: high-quality slices have stronger feature retention, while weak or noisy slices are controlled prior to decoder fusion.
Results and Discussion
Table 4 shows the quantitative results. The highest Dice, IoU, precision, recall, F1-score and accuracy were achieved by QGDL-Net. The Dice score of U-Net rose from 0.9471 to 0.9747 for QGDL-Net. The IoU for the model rose from 0.8995 to 0.9506, which means that the model has a higher degree of overlap with the expert mask and has fewer errors along the boundaries. QGDL-Net was further advanced than the U-Net with the quality-guided feature selection factor, while the attention mechanism also served to enhance performance over the U-Net. The relative improvement of QGDL-Net over the baseline models is reported in Table 5.
| Model | Dice | IoU | Prec. | Rec. | F1 | Acc. |
| U-Net | 0.9471 | 0.8995 | 0.9507 | 0.9435 | 0.9471 | 0.9963 |
| Attention U-Net | 0.9629 | 0.9285 | 0.9664 | 0.9594 | 0.9629 | 0.9974 |
| QGDL-Net | 0.9747 | 0.9506 | 0.9755 | 0.9739 | 0.9747 | 0.9982 |
Table 4. Performance comparison.
| Comparison | Dice gain | IoU gain | Interpretation |
| QGDL vs U-Net | +0.0276 | +0.0511 | Better overlap |
| QGDL vs Att-U-Net | +0.0118 | +0.0221 | Added quality value |
| Visual | Cleaner | Focused | Less leakage |
Table 5. Improvement summary.
The most significant gain over U-Net is achieved in IoU with an improvement of +0.0511 for QGDL-Net. This gain implies that the quality gate benefits the spatial conformity of the entire forecasted area, as IoU penalizes missing tumor pixels as well as extra predicted foreground. The precision value of 0.9755 also shows that QGDL-Net decreases background leakage from the baselines.
Although this is less significant than the improvement of Attention U-Net, it is still noticeable. There is no great concern for a significant performance gap since spatial gating is already used by Attention U-Net. The results of the additional measures, Dice gain (+0.0118) and IoU gain (+0.0221), however, indicate that channel regulation using quality information also contributes to the spatial attention. This confirms the hypothesis of a difference between the reliability of features and the relevance of their spatiality in segmentation.
The improvement in recall from 0.9435 to 0.9739 indicates that the QGDL-Net performs better in terms of recall compared to the U-Net. The reason for this is that under-segmentation may result in a reduced estimated tumor volume. The precision gain demonstrates that the proposed method is able to control false-positive regions, which is important when tumor regions are adjacent to normal hyperintense tissues. The Dice coefficient and the F1 score of QGDL-Net are equal because they both rely on the same foreground overlap formulation and are derived for binary segmentation. For all images, the accuracy values are high since most of the image pixels are background. So, just accuracy isn’t enough to prove that it’s the best; it’s the combined improvement in Dice, IoU, precision and recall.
The performance of QGDL-Net is 2.91% and 5.68% better than U-Net in terms of Dice score and IoU, respectively. The proposed model achieves a Dice metric improvement of 1.23% and an IoU metric improvement of 2.38% against Attention U-Net. While these differences are numerically small, they can make a significant visual difference in the boundary accuracy in challenging slices when measuring in segmentation.
The segmentation masks obtained in the proposed method are more compact and better aligned with the ground truth as shown in the segmentation overlays of Figure 3. This is especially noticeable for small and irregular tumor areas with only one baseline U-Net, where over-segmentation and under-segmentation occur. The improvement is in line with the way the quality gate is supposed to work: it prevents weak background activations from reaching the skip fusion stage.

Figure 3. Segmentation comparison for representative validation slices.
These visual differences are important for clinical decision-making, for instance, to calculate tumor burden from segmentation maps or for directing subsequent radiomic feature extraction. A mask that extends into normal tissue can affect volume and texture features, and a mask that is not completely around the tumor can lead to underestimation of tumor dimensions. The proposed model minimizes both errors for the representative samples. The visual comparison also shows an important practical benefit of the quality gating. Tumor texture is bright but irregular in slices where texture is present; baseline predictions could fragment or spill over into neighbouring tissue. QGDL-Net creates smoother and more anatomically consistent masks as it is a decoder that gets recalibrated features, not raw encoder features. This enhances visual coherence without the need for post-processing.
QGDL-Net only introduces small fully connected Quality Gates as an additional layer to the baseline encoder from a computational standpoint. Major evolutionary changes to the convolutional pathway are still U-Net-style, which makes the proposed model easier to train than large transformer models.
High-response areas are concentrated around the tumor-bearing areas, as seen in Figure 4 using EigenCAM visualization. EigenCAM is not a validation process, but it offers a qualitative assessment that QGDL-Net utilizes lesion-relevant regions. The metric plot (Figure 5) also indicates that the performance does not improve solely with one metric, but it is seen to improve across all the metrics that measure overlap and pixel-classification.

Figure 4. EigenCAM explainability map of QGDL-Net.

Figure 5. Graph plot comparing Dice, IoU, precision, recall, F1 and accuracy.
The same can be seen from the confusion matrix and the ROC curve of the experiment too. The predicted probability maps result in good foreground-background separation, and the confusion matrices reveal fewer FP and FN errors for the proposed model. These results match the numerical results on precision and recall improvement.
One limitation of this study is that only binary FLAIR-based segmentation was assessed. A multimodal and multiclass BraTS segmentation would give a more robust clinical experiment. Furthermore, the current validation split is an internal one, so an external test should be conducted between the institution and scanners prior to deployment. The controlled comparison proves that quality-guided gating can help improve a U-Net-style architecture even with a smaller transformer backbone.
Comparative Interpretation
A performance improvement is observed when comparing the performance of U-Net, Attention U-Net and QGDL-Net. The encoder-decoder baseline is given by U-Net. This is enhanced by learning spatial relevance for skip connections, which is done by an attention U-Net. QGDL-Net further enhances the performance by making image quality dependent on feature transfer. Thus, there is a logical experimental progression, not a random choice of baseline models, in the three-model comparison.
QGDL-Net is not a substitute for attention or transformer research but an orthogonal idea. Future systems could merge spatial attention with the transformer context and quality-governed gates into a single network. In the present work, the quality component is separated so its effect is clear. Isolation is good for methodological papers because it minimises the problem of attributing improvement to more than one unchecked change.
Practical Implications
For radiology support tools, robustness is as crucial as peak accuracy. A model which is effective for high-contrast slices will not work in subtle tumor areas. Through feature transfer based on quality statistics, QGDL-Net prompts the network to adapt to the reliability level at the slice level. This is especially helpful in data sets obtained from various scanners, protocols and institutions.
The explainability result also lends credence to the interpretability. Eigen-CAM maps are focused around tumor regions and overlap with the general location of predicted maps. This isn’t a clinical correctness check, but rather a way to check if the network is using plausible image regions. This visualization, together with overlays on the segments and quantitative measurements, can increase confidence in the model output.
Conclusion
This paper introduced QGDL-Net for brain tumor magnetic resonance imaging (MRI) segmentation in the context of quality control. The proposed method performs encoder feature recalibration prior to decoder fusion with the help of a slice-level quality descriptor and learnable channel gates. Using mathematical analysis, it was found that sigmoid quality gates were unable to boost the energy of noisy features, and experiments conducted on BraTS 2021 yielded better performance than U-Net and Attention U-Net. QGDL-Net achieved a Dice of 0.9747 and an IoU of 0.9506. The method will be expanded to multimodal 3D segmentation of subregions, external validation and uncertainty-aware clinical deployment (future work).
The findings indicate that there is still a role for compact reliability-aware modules in today’s age of large transformer and hybrid networks. QGDL-Net does not add depth but enhances the usefulness of the information already available in the encoder features. The method is appealing for resource-poor experiments and for applications where simplicity of computation and interpretability are crucial.
Future direction is uncertainty estimation. A clinically useful segmentation system ought to not only generate a mask but also convey confidence. A combination of Quality gates with Monte Carlo dropout and deep ensembles/calibration losses can be used to detect uncertain borders. These additions would transform the network to be more radiotherapy planning and radiomic-friendly.
Another implementation extension is to use a small quality vector instead of a scalar quality value, consisting of contrast, entropy, edge strength, tumor-area prior and local intensity variation. Such a vector would enable the gate to discriminate between various causes of poor segmentation, such as low contrast, small lesion size and noisy background tissue. In clinical translation, the predicted mask is to be evaluated along with the radiologist’s review. QGDL-Net can be used as a first approximation for contour generation and then corrected by hand. This would save time for annotating the data, but at the same time provide expert supervision, an essential condition for safety-critical applications like neuro-oncology.
References
- [1]Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., et al.: The multimodal brain tumor image segmentation benchmark (BRATS). IEEE Transactions on Medical Imaging 34(10), 1993–2024 (2015)DOI
- [2]Bakas, S., Akbari, H., Sotiras, A., Bil ello, M., Rozycki, M., Kirby, J.S., et al.: Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data 4, 170117 (2017)
- [3]Baid, U., Ghodasara, S., Mohan, S., Bil ello, M., Calabrese, E., Colak, E., et al.: The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314 (2021)
- [4]Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomedical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), pp. 234–241. Springer International Publishing, Cham (2015)
- [5]Çiçek, Ö., Abdulkadir, A., Lienkamp, S., Brox, T., Ronneberger, O.: 3D U-Net: Learning dense volumetric segmentation from sparse annotation. In: Medical Image Computing and Computer-Assisted Intervention, pp. 424–432 (2016)
- [6]Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnU-Net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 18, 203–211 (2021)
- [7]Oktay, O., Schlemper, L., Folgoc, L., Lee, M., Heinrich, M., Misawa, K., et al.: Attention U-Net: Learning where to look for the pancreas. In: Medical Imaging with Deep Learning (2018)
- [8]Wang, W., Chen, C., Ding, M., Yu, J., Zha, H., Li, J.: TransBTS: Multimodal brain tumor segmentation using transformer. Medical Image Analysis 84, 102678 (2023)
- [9]Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., et al.: UNETR: Transformers for 3D medical image segmentation. In: IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 574–584 (2022)
- [10]Jiang, Y., Zhang, Y., Lin, X., Dong, J., Cheng, T., Liang, J.: SwinBTS: A method for 3D multimodal brain tumor segmentation using swin transformer. Frontiers in Oncology 12 (2022)
- [11]Zhou, H., Guo, J., Zhang, Y., Yu, L., Wang, L., Yu, Y.: nnFormer: Volumetric medical image segmentation via a 3D transformer. IEEE Transactions on Image Processing 32, 4036–4045 (2023)
- [12]Zhang, W., Chen, S., Ma, Y., Liu, Y., Cao, X.: ETUNet: Exploring efficient transformer enhanced UNet for 3D brain tumor segmentation. Computers in Biology and Medicine 171, 108005 (2024)
- [13]Zeineldin, R.A., Karar, M.E., Mathis-Ullrich, F., Burgert, O.: Explainable hybrid vision transformers and convolutional network for multimodal glioma segmentation in brain MRI. Scientific Reports 14 (2024)
- [14]Preetha, R., M., J.P.P., S., N.J.: Brain tumor segmentation using multi-scale attention U-Net with EfficientNetB4 encoder for enhanced MRI analysis. Scientific Reports 15, 9914 (2025)
- [15]Shahid, G.E.S., Hossain, M.A., Rahman, S., Islam, M.M., Latif, S.A.: LIU-NET: Lightweight Inception U-Net for efficient brain tumor segmentation from multi-modal 3D MRI images. PeerJ Computer Science 11, e2787 (2025)
- [16]Tabassum, M., Di Ieva, A., Liu, S.: Meta transfer learning for brain tumor segmentation using nnUNet in meningioma and metastasis cases. Scientific Reports 15(1), 37599 (2025)