A patch-wise deep residual network (PwDRU-Net102) for multimodal MRI brain tumor segmentation

1 Introduction

The high rates of illness and death associated with brain tumors make them a major worldwide health concern. According to the International Association of Cancer Registries (IACR), about 24,000 people die from cancer in India each year, with over 28,000 new cases that are identified (Dasgupta et al., 2016; Liu et al., 2022). In the United States, brain tumors accounted for more than 13,000 deaths in 2019, with nearly 87,000 diagnosed cases (Patel et al., 2019; Cancer Research, 2019). Benign (noncancerous) or malignant (cancerous) tumors are caused by the uncontrolled development of aberrant brain cells (Lapointe et al., 2018). The most frequent primary brain tumors are gliomas, which are derived from glial cells. They are further divided into two types: slow-growing Low-Grade Gliomas (LGG) and aggressive-behaving High-Grade Gliomas (HGG).

Brain tumors are further divided into two types: primary, where the tumors originate inside the brain, and secondary, which are metastases from other organs like the skin, breasts, intestines, kidneys, or lungs. High-grade tumors require immediate attention. They are typically diagnosed using imaging, surgery, or cerebrospinal fluid tests (Louis et al., 2016). These conventional diagnostic methods involve risks, such as bleeding, infection, and limited diagnostic precision, inherent in procedures like biopsy and fluid aspiration (Kasraeian et al., 2010). Medical imaging plays an important role in tumor grading and detection. Among various modalities, such as CT, X-rays, ultrasound, MRI is one of the favorable techniques as it is non-invasive and offers complete tissue contrast (Al-Qazzaz, 2020; Ghasemi et al., 2025). Various MRI sequences, such as FLAIR, T1-weighted (T1w), T2-weighted (T2w), and T1 contrast-enhanced (T1ce), emphasize complementary features of tissues (Baid et al., 2020; Dehghani et al., 2024). Given this purpose of treatment, exact tumor segmentation from these images becomes necessary. Automated techniques offer uniformity and efficiency; however, manual segmentation is very time-consuming and requires special expertise. Besides all these improvements, irregular shapes and heterogeneous appearance of tumors pose challenges to segmentation algorithms (Jyothi and Singh, 2023).

Recent developments in machine learning and deep learning have made it possible to identify, classify, and segment brain tumors automatically (Dorfner et al., 2025; Xiao et al., 2025a). In contrast to deep learning, especially Convolutional Neural Networks (CNNs), which can automatically extract pertinent features but frequently require large-scale datasets, classical machine learning depends on handmade features and smaller datasets (Hollon et al., 2020; Ranjbarzadeh et al., 2021; Verma and Yadav, 2025). Traditional patch-based segmentation algorithms process small image regions in isolation, often leading to the loss of global contextual information, boundary discontinuities, and fragmented tumor predictions. Such techniques struggle to capture long-range spatial dependencies across an entire MRI slice due to their limited receptive fields.

To address these problems, a Patch-wise Deep Residual U-Net architecture with 102 convolutional layers i.e., PwDRU-Net102 has been proposed in this study by adopting block-based processing in conjunction with a deep residual U-Net architecture. This design enables effective integration of local feature representations with global contextual information, resulting in more accurate and coherent tumor segmentation.

The model partitions MRI scans into non-overlapping patches, preserving local features while using residual blocks to prevent vanishing gradient issues. Dropout layers are employed to reduce overfitting. Based on experimental evaluation, PwDRU-Net102 provides an accurate and reliable framework, soutperforming current methods in terms of Hausdorff Distance (HD) and Dice Similarity Coefficient (DSC).

Main contributions of this study are as listed.

• For brain tumor segmentation, a robust PwDRU-Net102 model with 102 convolutional layers was trained using the BraTS (2019 and 2020) MRI datasets.
• Incorporation of uniform, non-overlapping patch-wise MRI slices to improve local feature learning and tumor localization.
• Dropout layers are used in the encoder to address the overfitting issue after every stack of convolutional blocks and the bottleneck residual block.
• Employment of bottleneck residual blocks with shortcut connections to prevent vanishing gradient problems.
• Deployed a high number of training data to improve the model’s performance.
• Demonstrated improved segmentation performance compared over selected existing methods, evaluated with DSC and HD metrics.

The following is the structure of the paper: State-of-the-art techniques have been discussed in Section 2, Section 3 details the proposed 2D Patch-wise Deep Residual U-Net102 approach, Section 4 presents the results of the experiments and statistical analysis. Section 5 provides a comparative analysis, and Section 6 concludes the study and outlines future research directions.

2 Literature survey

Brain tumors are life-threatening conditions that demand timely detection and treatment. In order to diagnose and treat brain cancers, MRI scans must be precisely segmented. Research in this area has advanced from traditional machine learning to deep learning models.

One of the early approaches was traditional machine learning, which involved manually engineered features (Bauer et al., 2013). Li et al. (2019) introduced a fuzzy C-means framework on the BLSA dataset to improve temporal consistency, while Saif et al. (2019) employed whole-brain tractography with one-class SVM to isolate brain tracts. Sun et al. (2019) proposed the mcRBM model to identify regions of interest in MRI data. Although adaptable, such methods are computationally expensive and often underperform on high-dimensional and imbalanced datasets.

Brain MRI segmentation was greatly enhanced by the development of CNN-based deep learning. Sliding-window CNN with multi-scale patches was designed by Moeskops et al. (2016), and Nie et al. (2016) applied fully convolutional networks (FCNs) for multi-modal MR images. Wachinger et al. (2018) have applied CNNs for neuroanatomical segmentation, while Bao and Chung (2018) further extended them with multi-scale and dynamic random walker refinements. Multi-model approaches were introduced by Wei et al. (2019) that combined U-Nets with autoencoders and probabilistic atlases. Milletari et al. (2017) proposed the Hough-CNN for automatic localization. Still, the patch-based CNN methods are computationally heavy and sometimes lack sufficient context.

Encoder-decoder models, especially U-Net (Ronneberger et al., 2015), have achieved significant improvements due to the saving of fine-level features through skip connections. Several variations have been proposed afterwards (Abd-Ellah et al., 2019; Zhang et al., 2020a; Maji et al., 2022; Qamar et al., 2020; Isensee et al., 2020; Huang et al., 2021). However, the SegNet (Badrinarayanan et al., 2017) and U-Net (Ghaffari et al., 2020) variants still have their drawbacks regarding slower training or instability in the gradients. In addition, to handle these issues, several residual architectures, such as ResNet (He et al., 2016a; He et al., 2016b) and DeepLabV3+ (Chen et al., 2017; Chen L. C. et al., 2018), introduced identity mappings, bottleneck blocks, and spatial pyramid pooling for deeper feature learning. Regularization techniques like dropout (Park and Kwak, 2016) helped reduce overfitting. It is still challenging to manage the appearance, dimension, and structure diversity of gliomas (Azhari et al., 2014). Xiao et al. (2025b) introduced an advanced U-Net model that provides improved feature learning and feature fusion capabilities using residual connections and attention mechanisms demonstrating better segmentation results.

A thorough review of the existing literature on U-Net, residual U-Net, and patch-based CNN models indicates that, while these approaches are well established and effective, each suffers from specific limitations. Motivated by these observations, we propose a hybrid model that capitalizes on the strengths of all three approaches. Specifically, we introduce a Patch-wise Deep Residual U-Net (PwDRU-Net102), which integrates uniform non-overlapping patches and residual blocks within a 102-layer encoder–decoder framework to enhance training stability, localization capability, and segmentation accuracy. By combining residual connections with patch-wise learning in a U-Net architecture, the proposed method addresses the identified shortcomings of existing approaches and aims to achieve more stable and accurate segmentation. A concise comparison of prior methods, their limitations, and how the proposed approach addresses them is summarized in Table 1.

TABLE 1. Comparison of previous methods, their limitations, and the design choices through which the proposed PwDRU-Net102 overcomes these limitations.

3 Methodology

The proposed structure and step-by-step procedure of the study have been discussed with a schematic representation shown in Figure 1. Training and testing have been done using MRI scans from the BraTS datasets (Bakas et al., 2018). To remove unnecessary signals, noise, and artifacts, the data was preprocessed (Suhas and Venugopal, 2017) before the model was developed. Five techniques were applied during pre-processing: data transformation, subset division, narrow region filtering, morphological operations, and feature scaling.

Following preprocessing, the proposed segmentation model, PwDRU-Net102, was then implemented. Each MRI slice of size 256 × 256 was partitioned into four non-overlapping patches of 128 × 128. To minimize computational redundancy and memory overhead, non-overlapping patches are employed. Compared to overlapping patches, this strategy avoids repeated processing of the same regions and improves training efficiency, particularly for high-resolution images. While this approach promotes stable convergence during training, it may introduce minor boundary effects, which are addressed through residual feature propagation in the proposed architecture.

In this context, patch-based segmentation methods employing overlapping patches have been shown to capture boundary details more effectively than non-overlapping approaches, owing to the additional contextual information available at the boundaries of neighbouring regions. However, this advantage is often accompanied by increased computational cost due to repeated processing of shared pixels across overlapping areas, leading to longer training and inference times as well as higher memory access overheads (Liu et al., 2018). In contrast, full-image training inherently preserves global spatial context and long-range dependencies within an image. Nevertheless, for high-resolution medical images, its applicability is frequently limited by GPU memory constraints, rendering such approaches impractical for certain deep architectures (Jeon et al., 2025).

Therefore, the proposed work adopts a non-overlapping patch-based strategy as a balanced compromise. This approach reduces computational redundancy and memory overhead while still achieving competitive segmentation performance when compared to overlapping patch-based methods and full-image training approaches.

In addition to the patching strategy, the choice to use 102 convolutional layers in the proposed PwDRU-Net102 architecture was guided by both theoretical considerations and experimental validation. Increasing the depth of deep neural networks has been shown to enhance the extraction of hierarchical and complex features, which is critical for accurate brain tumor segmentation. However, excessively deep architectures often suffer from issues such as vanishing gradients and overfitting.

To address these challenges, the proposed model incorporates residual learning modules, which facilitate effective gradient flow and improve training stability in deeper networks. From a theoretical perspective, residual connections enable identity mapping and help preserve feature representations across layers, allowing the network to benefit from increased depth without degradation in performance. Additionally, the use of dropout mitigates overfitting by preventing the co-adaptation of neurons. From an experimental standpoint, architectures with varying depths (e.g., 50, 75, and 102 layers) were evaluated during model development. The results demonstrated that the 102-layer configuration provided the optimal trade-off between segmentation accuracy and computational efficiency. Although deeper architectures resulted in slightly increased training time, they achieved significantly improved feature representation and more precise tumor boundary detection, leading to higher Dice similarity scores. Further increases in depth beyond 102 layers yielded only marginal performance gains while incurring substantially higher computational costs.

The proposed method PwDRU-Net102 consists of 102 convolutional layers with encoder–decoder paths, residual blocks, and adaptive skip connections. Dropout regularization was introduced at varying rates (0.0, 0.2, and 0.5) across layers to reduce overfitting. Residual blocks further alleviated vanishing gradient problems, ensured smoother feature transmission, and supported faster model convergence. Both input patches and corresponding ground-truth masks were used during training and testing.

3.1 Pre-processing

In medical image analysis, pre-processing is a crucial initial step for accurate and consistent results.

The BraTS (2019 and 2020) datasets were already partially preprocessed by the competition organizers, including isotropic resolution, skull stripping, co-registration, and alignment to a common space (Menze et al., 2015). However, MRI scans from different scanners often exhibit intensity variations, which can affect segmentation accuracy. Therefore, additional preprocessing was performed to remove artifacts and standardize the data (Tambe et al., 2013; Patro and Sahu, 2015).

• Data transformation: MRI scans come in all sorts of formats, but perhaps the most common format is that of NIfTI (.nii.gz) files. The images were imported using Python 3.9 with neuroimaging libraries Nibabel and Nilearn and converted to Numpy arrays to allow for faster computation and reduction in memory usage. Nilearn was further used for visualization. The goal of this transformation was to provide a uniform structure across all datasets.
• Subset division: In order to handle large volumes of data, the dataset was separated into HGG and LGG subsets. BraTS 2020 consists of 1,675 patient scans (1,173 HGG, 502 LGG), BraTS 2019 consists of 2,970 scans (2,079 HGG, 891 LGG), and the ground-truth maps were correspondingly extracted. Data were then shuffled and divided into training and testing components in a 70:30 ratio to enable an unbiased evaluation of the model.
• Narrow region filtering: It is a filtering approach that concentrates on specific regions of pictures, such as high intensity or aberrant texture, where important MRI scan characteristics (like tumors) emerge. Eliminating irrelevant regions is the primary goal in order to lower false positives.
• Feature scaling: Z-score normalization was applied to standardize feature ranges, centering data at mean 0 and standard deviation 1 (Patro and Sahu, 2015). This improves learning stability, reduces sensitivity to outliers, and helps the model converge more effectively.

After preprocessing, clean and standardized MRI scans were divided into uniform, non-overlapping patches capturing both tumor and surrounding brain tissue. Following that, the DRU-Net102 model was fed these patches as input for segmentation.

3.2 Segmentation

Due to its ability to extract both local and global information through skip connections between the encoder and decoder layers, U-Net is a widely used architecture for image segmentation. The model can produce segmentation masks by integrating high-level and low-level information due to these connections. Despite its advantages (Akram and Usman, 2011; Roshanzadeh and Afrakhteh, 2019; Arora et al., 2021), U-Net has certain limitations that affect performance and generalization (Dhage et al., 2015; AS and Gopalan, 2022).

In deep networks, the gradients in the early layers can become extremely small during backpropagation, which slows or even halts learning in those layers. This leads to the loss of crucial low-level spatial details (edges, boundaries, location cues) during downsampling, even though skip connections (Yang and Yang, 2018) transfer features across layers. Second, the model is prone to overfitting, as it may adapt too closely to noisy or imbalanced training data, reducing its ability to generalize to unseen samples. Other challenges include dead neurons in activation and difficulties with convergence in very deep architectures.

For brain tumor segmentation, this paper presents the Patch-wise Deep Residual U-Net102 (PwDRU-Net102) in order to overcome these problems. The model incorporates the encoder-decoder architecture of U-Net with residual blocks (Srivastava et al., 2015; Messaoudi et al., 2020) and identity mappings to retain low-level features, ensure smoother information flow, and mitigate vanishing gradients. Shortcut connections directly pass activations from earlier to later layers, improving stability and convergence.

Additionally, dropout regularization (Ballestar and Vilaplana, 2020) was incorporated after each level with varying rates (0, 0.2, 0.5). Experimental results showed that a dropout rate of 0.2 achieved the best trade-off, reducing overfitting while maintaining strong performance (Park and Kwak, 2016). PwDRU-Net102 with deeper residual learning and patch-wise training enhances generalization compared to the conventional U-Net.

3.2.1 Proposed model

The MRI scans, now preprocessed, contained only relevant information and were ready for the segmentation process. However, before segmentation, each MRI image was divided into four uniform, non-overlapping patches. This patch-wise approach plays a crucial role in preserving local features. Moreover, since patches require less memory during computation, they enable faster training compared to using entire images.

The BraTS dataset initially consisted of MRI volumes, each having a size of 240 × 240 × 155, with three modalities: T1, T2, and FLAIR. For each scan, about 48 axial slices were selected. To standardize input, zero-padding was used to expand the slices to 256 × 256, adding 16 pixels vertically and 40 horizontally, which resulted in an input volume size of 256 × 256 × 48. These modalities capture variations in tissue and cerebrospinal fluid intensities that are necessary for the identification of tumors. For model training, each 256 × 256 size image and its associated ground-truth segmentation image were split up into four 128 × 128 image patches for training. Both training and testing of the PwDRU-Net102 model were conducted using these patches. During inference, the predicted segmentation maps of individual blocks are reassembled to reconstruct the full 256 × 256 segmentation output. During reconstruction, contextual features preserved within each block ensure continuity across block boundaries. In contrast to small patch-based approaches, which often produce sharp and discontinuous edges due to limited patch size, the larger block-based strategy, combined with residual feature propagation, effectively smooths block boundaries and reduces edge artifacts. All modalities were used to assess segmentation performance on the three tumor classes: WT (Whole Tumor), ET (Enhancing Tumor) and TC (Tumor Core). By using patch-wise inputs, the model achieves improved localization, better feature representation, and faster convergence.

3.2.2 PwDRU-Net102 architecture

The four key components of the proposed architecture are the bottleneck, decoder, skip connections, and encoder. This design uses distinct stacked residual block stages including layers of convolution and identity mapping for the encoder and decoder paths, which are connected by the skip component.

As illustrated in Figure 2, each patch slice of a 2D MRI image discussed above has been fed to the proposed architecture as the input. The residual block includes batch normalization, the Rectified Linear Unit (ReLU) activation function, and two or three convolutional layers in addition to accepting the input layer. To extract visual information like edges and patterns, each convolutional layer employs a tiny (3 × 3) filter. Batch normalization is a method for normalizing the layer’s output and preventing vanishing gradients. It improves the speed, depth, and stability of data training. The activation function known as the ReLU enables the model to learn complicated and non-linear patterns. In order to make feature mapping easier, the input of each decoder level has been added to the encoder’s output (after upsampling) using a concatenation link. When utilized in deep networking, it is reliable and efficient. Figure 3 illustrates how Batch Normalization has been implemented in each residual block, following convolution and preceding ReLU.

The encoder is a feature extractor that captures various levels of patterns, edges, and textures while extracting features step-by-step. Each layer goes deeper, comprehending more intricate characteristics such as the borders or forms of tumors. In this process, residual blocks facilitate learning and help retain crucial information. The result is a tiny, abstract image that contains crucial information known as encoded features. The decoder uses the encoded information to reconstruct the segmented image (tumor vs. non-tumor). The little feature maps are enlarged back to their original size using upsampling or transpose convolution in the decoder path. Between the encoder and decoder, the U-Net architecture includes a bottleneck block, which forms the deepest layer and serves as a bridge between the encoder and the decoder. This block employs residual convolutional layers to learn high-level semantic features over a large receptive field. Residual connections ensure smooth feature and gradient flow throughout the network, facilitating stable training of deep architectures. The bottleneck block effectively fuses global contextual information before upsampling in the decoder stage. This design enhances feature transition between the encoder and decoder, improves training convergence, and ultimately leads to more accurate segmentation results.

Finally, a segmentation mask that identifies the tumorous and non-tumorous areas of the image is produced by the decoder. By using skip connections, crucial low-level information can be sent straight from the encoder level to the corresponding decoder level, which aids in avoiding vanishing gradients. Afterwards, we have implemented different dropout regularizers (0.2, 0.5 and 0) after each level. In the first convolution of each level, downsampling has been executed with a stride of two (except for level 1) which reduces spatial dimensions. With kernel size = 2 × 2 and stride = 2, transposed convolution was used in the decoder to upsample the image while restoring its size. Pixel-wise classification scores are derived using a 1 × 1 kernel convolving with a softmax activation function to represent the tumor class in the final layer.

The detailed parameter setup of the suggested approach is demonstrated in Table 2. Tables 3–5 illustrate the final segmentation map produced by the proposed approach in comparison to the ground truth segmentation map using the stack of T1, T2 and FLAIR MRI sequences respectively.

TABLE 2. Parameter configuration of PwDRU-Net102.

TABLE 3. Segmentation results obtained using the proposed PwDRU-Net102 for T1-weighted MRI sequences.

TABLE 4. Segmentation results obtained using the proposed PwDRU-Net102 for T2-weighted MRI sequences.

TABLE 5. Segmentation results obtained using the proposed PwDRU-Net102 for FLAIR MRI sequences.

The PwDRU-Net102 model generates a pixel-level classification map as its final output, which is saved in the database for evaluation and testing. The ground truth mask is compared to the predicted segmentation map in order to evaluate the accuracy. Unseen MRI brain scans are assessed against the trained database to classify each case as benign (0) or malignant (1) during the testing phase where we have applied Artificial Neural Network (ANN) (Shakya, 2020) and SVM (Azarmdel et al., 2020) for Brain Tumor Classification (Usman and Rajpoot, 2017; Raza and Singh, 2021; Jeyaraj and Samuel, 2019; Cai et al., 2020). The final feature vectors obtained after the implementation of ANN and SVM, MRI scans can be classified into normal and malignant brain scans.

4 Evaluation analysis

4.1 Data gathering

The dataset used to examine brain lesions was obtained from the “University of Pennsylvania, Center for Biomedical Image Computing and Analysis” (CBICA) and the Multimodal BraTS challenge (Lachinov et al., 2018). Using the publicly accessible BraTS benchmark datasets (2019 and 2020) challenges for segmentation and classification of MRI brain tumor, the proposed method has been trained and evaluated. To prevent data leakage, the dataset was split at the patient level, such that all slices from a given patient were included exclusively in either the training set or the testing set. Here we have applied Data Augmentation technique to achieve significant variation to the data. After Data augmentation, a total of 2970 MRI brain images were utilized in the BraTS 2019 dataset, of which 2079 were from HGG and 891 were from LGG. A total of 1675 MRI brain images were used in the BraTS 2020 dataset, 1,173 of which were HGG and 502 were from LGG. This dataset has been applied to a standard 256 × 256 MRI scans for model training as well as testing. Four MRI image sequences with the matching segmentation results have been included in each patient’s folder in the BraTS dataset. MRI images (T1, T2 and FLAIR weighted) have been employed for each dataset, and our suggested method applied these sequences separately.

4.2 Statistical parameters

Two commonly used metrics for analyzing the performance of segmentation techniques, the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD), have been utilized to evaluate the efficiency of the suggested approach (Menze et al., 2015). The degree of agreement between the related ground truth segmentation map and the anticipated segmentation map is measured by them. The overlap of the ground truth and predicted segmentation scans have been determined by the DSC. Effective segmentation is indicated by a higher DSC value, which ranges from 0 to 1.

The tumor structure is split into three sub-regions in order to evaluate the model’s potential for biological applications. The first is Enhancing Tumor (ET), which usually denotes the High-Grade Glioma (HGG) regions that are actively enhancing. Subsequently, the Tumor Core (TC) encompasses every component of the tumor except for the edema, and the Whole Tumor (WT) comprises all the malignant regions. The label for tumor voxels is 1, while the label for non-tumor voxels is 0 for both the ground truth and the predicted segmentations.

A boundary-based metric called HD95 calculates the distance between the borders of two segmentation masks i.e., predicted and ground truth. It reduces sensitivity to outliers by ignoring the worst 5% of errors and instead uses the 95th percentile of the distances, rather than the maximum. An HD95 value of zero indicates a perfect boundary match, whereas higher HD95 values reflect poorer boundary alignment. Furthermore, this study examined the Mean Squared Error (MSE). The MSE is a representation of the average squared differences between the segmentation map of the ground truth and prediction. Other parameters, namely, Accuracy (Acc), Specificity (Sp), and Sensitivity (Sen), have also been examined as part of the statistical analysis to derive discrete evaluation values based on the confusion matrix.

4.3 Result and discussion

Here, this section highlights the experimental analysis of the different models DRU-Net104, DRU-Net and DRU-Net102 with our proposed model (PwDRU-Net 102) using above stated statistical parameters. DRU-Net104 (Colman et al., 2021) is a deep residual U-Net architecture comprising 104 convolutional layers. It follows the standard encoder–decoder U-Net design, with residual blocks incorporated at every level. In contrast to the proposed PwDRU-Net102, DRU-Net104 does not employ block-level (patch-wise) processing and instead operates directly on full-resolution input images. Moreover, the block-based segmentation strategy that forms a key novelty of PwDRU-Net102 is not present in DRU-Net104.

The DSC (ET, WT, and TC) outcomes of the above models were examined using the 2019 and 2020 BraTS datasets’ numerical findings, which are shown in Tables 6 and 7, respectively. Here, DSC analysis using various algorithms as mentioned above has been performed on three tumor regions, namely, ET, TC, and WT. For WT segmentation on the BraTS 2020 dataset, the proposed method outperformed the other algorithms with a Dice score of 91.36% when compared to ground truth.

TABLE 6. Dice Similarity Coefficient (DSC) results for the BraTS 2019 dataset.

TABLE 7. Dice Similarity Coefficient (DSC) results for the BraTS 2020 dataset.

Tables 8 and 9 present a comparison of the proposed method for numerical analysis of HD95 with DR-Unet104, DR-Unet and dRU-Net102 for stated MRI images using BraTS 2019 and 2020 datasets, respectively. Here, HD95 analysis has been done using TC, WT and ET tumor regions. A lower HD95 of BraTs 2020 of WT tumor region (15.78) indicates improved boundary alignment.

TABLE 8. 95th percentile Hausdorff Distance (HD95) results for the BraTS 2019 dataset.

TABLE 9. 95th percentile Hausdorff Distance (HD95) results for the BraTS 2020 dataset.

To solve the problem of overfitting, we have applied DSC with different dropouts (0, 0.2 and 0.5) on our proposed algorithm i.e., PwDRU-Net102 using BraTs 2019 and BraTS 2020 datasets as depicted in Table 10 and Table 11 respectively. The comparative analysis of different dropouts shows that at dropout 0.2, result of WT (0.9081) of BraTS 2020 exhibits significant overlapping between ground truth and the prediction. It means that our proposed model performs exceptionally well when it comes to segmenting the entire tumor region.

TABLE 10. Dice Similarity Coefficient (DSC) results with dropout for the BraTS 2019 dataset.

TABLE 11. Dice Similarity Coefficient (DSC) results with dropout for the BraTS 2020 dataset.

Table 12 differentiates the proposed approach with DRU-Net, DRU-Net104, and DRU-Net102 in terms of numerical investigation of Accuracy for all stated MRI scans employing BraTS datasets. Table 12 illustrates the 99.54% accuracy of our proposed technique for T1 MRI sequences (BraTs 2020). The different model accuracy changes over various training epochs are depicted in the graph (Figure 4). Our proposed method (green line) reaches a peak accuracy close to 1.0. It achieves high accuracy within just 2-3 training epochs, which shows the fastest convergence. The model outperforms all others in both accuracy and computational efficiency.

TABLE 12. Comparative accuracy analysis of different segmentation methods on BraTS 2019 and 2020 datasets across T1, T2, and FLAIR MRI sequences.

Using BraTS 2019 and 2020 datasets, Table 13 illustrates Specificity numerical results. The table compares the proposed PwDRU-Net102 model with DRU-Net, DRU-Net104, and DRU-Net102, highlighting that our model achieves a specificity of 99%.

TABLE 13. Analysis of specificity.

Table 14 presents a numerical results for Sensitivity across all stated MRI scans, evaluating the proposed method against DR-Unet, DR-Unet104, and DRU-Net102 using the 2019 and 2020 editions of the BraTS dataset. A sensitivity of 97.92% was attained using the recommended approach.

TABLE 14. Analysis of sensitivity.

Using the confusion matrix of BraTS 2020 (Table 15) and BraTS 2019 (Table 16), with HGG labeled as Positive and LGG as Negative, the effectiveness of the suggested approach has been evaluated in comparison to dRU-Net102.

TABLE 15. Comparison of confusion matrices for PwDRU-Net102 and DRU-Net102 on the BraTS 2020 dataset.

TABLE 16. Comparison of confusion matrices for PwDRU-Net102 and DRU-Net102 on the BraTS 2019 dataset.

To address class imbalance and improve segmentation accuracy, the proposed PwDRU-Net102 model is trained using the Dice loss function, defined as (Pereira et al., 2016; Havaei et al., 2017):

L D i c e = 1 − 2 ∑ i p i g i + ∈ ∑ i p i + ∑ i g i + ∈

Where denotes the predicted probability of pixel i, represents the corresponding ground-truth label and is a small constant added to avoid division by zero.

The change in the loss value over training epochs of DRU-Net104, DRU-Net102 without patches and DRU-Net102 with patches (our model) has been plotted as shown in Figure 5. It demonstrates that how our suggested model gradually reduces the loss.

Figure 6 shows the graph of the various model accuracy changes over different validation epochs. Our suggested approach achieves the highest validation accuracy as shown in the figure with green line.

This graph (Figure 7) demonstrates that DRU-Net102 with patches is the best option out of the three models since it has the lowest and most consistent validation loss over time. This indicates that this model is the best at generalizing to new and unseen data.

The training of the proposed model involves loss function that was tailored to medical images segmentation. Training was done using a fixed number of epochs and Adam optimizer with a constant learning rate and a constant uniform batch size in all the experiments. The same GPU hardware was used to train and evaluate all the models in order to achieve consistency as shown in Table 17 below.

TABLE 17. Training Configuration of the proposed PwDRU-Net102 model.

In particular, among them, TC and ET show very challenging segmentation tasks in which traditional approaches often had lower accuracy, while PwDRU-Net102 handled such difficult regions effectively, as indicated by Table 18, with improved segmentation results compared to prior methods. Here 3D approaches has been included in the comparison table to give more context to the entire performance, and to also be able to point out methodological trade-offs, but not to imply that these approaches can be directly compared.

TABLE 18. Comparison of existing methods with our proposed approach on segmentation in terms of DSC.

5 Conclusion and future scope

Accurate brain tumor segmentation is required for prompt diagnosis and treatment. Our study addressed the challenges in medical image processing by proposing a new 2D Patch-wise Deep Residual U-Net with 102 convolutional layers, referred to as PwDRU-Net102. The architecture includes residual blocks with identity mappings in the encoder to maintain the local features that are transferred to the decoder through feature links, aiding in the reduction of vanishing gradients. Dropout regularization at multiple convolution stages further reduces overfitting.

The model has been evaluated with the BraTs Datasets of 2019 and 2020 with the aim of segmenting MRI brain tumors using T1, T2, and FLAIR sequences. Results show that PwDRU-Net102 achieves improved overlap and boundary accuracy for different regions in contrast to the previous approaches. Although it is a 2D approach and was trained using limited data, it performed well in generalization and gave reliable segmentation results.

While the proposed model demonstrates strong generalization performance, certain limitations should be acknowledged, particularly in relation to the non-overlapping block-based strategy. Segmentation performance may be affected for very small tumors due to the fixed block size. Minor boundary inconsistencies can arise when tumors are located near block edges or extend across multiple blocks, potentially influencing segmentation continuity. Furthermore, the fixed block approach may be less adaptable to tumors with highly irregular shapes or varying spatial extents. These factors indicate that tumor size and spatial location can impact segmentation performance. Consequently, further validation on additional datasets with diverse tumor characteristics is necessary to more comprehensively assess the generalization capability of the proposed method.

Patch-wise segmentation may not fully capture long-range dependencies across distant regions of the image. However, this design choice enables efficient training on high-resolution multimodal MRI while maintaining strong segmentation performance. Incorporating explicit global-context mechanisms will be explored in future work. Future work could extend the framework to 3D architecture and larger multimodal data to capture volumetric context more effectively. The model’s usefulness in practical medical applications can be further expanded by integrating it with clinical decision-support systems, which will help radiologists accurately and promptly assess tumors. Incorporating statistical significance testing in an extended version of this study.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Michael Dada, Federal University of Technology Minna, Nigeria

Reviewed by: Chaodong Fan, Hainan University, China

Vikash Verma, AISECT University, India