A hybrid deep learning framework based on VGG19 and U-Net for accurate brain tumor segmentation in MRI images

Abstract

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) is a critical task in neuro-oncology, directly impacting diagnosis, treatment planning, and patient outcomes. This study presents a hybrid deep learning model that integrates the VGG19 convolutional neural network as encoder and the U-Net architecture as decoder for automated tumor segmentation. The model is trained and validated using the lower-grade glioma (LGG) segmentation dataset from the cancer imaging archive (TCIA), which includes annotated FLAIR MRI scans of LGG cases. By leveraging the deep feature extraction capabilities of VGG19 and the spatial precision of U-Net, the proposed model achieves superior segmentation accuracy while maintaining architectural simplicity. Quantitative evaluation demonstrates a Dice of 92.20%, intersection over union (IoU) of 85.53%, and area under the curve (AUC) of 95.90%, outperforming several state-of-the-art models in the literature. This framework offers a reliable and efficient solution for clinical MRI analysis and has strong potential for integration into computer-aided diagnostic systems. The results indicate that hybrid convolutional neural network (CNN) architectures can significantly enhance the accuracy and robustness of brain tumor segmentation tasks. The novelty of this work lies in the seamless integration of VGG19’s pre-trained deep features with U-Net’s skip connections, providing a lightweight yet highly accurate model that reduces computational complexity compared to multi-branch or attention-based architectures, while achieving superior performance on the LGG dataset.

Graphical abstract

1. Introduction

Brain tumors represent a class of life-threatening neurological disorders that account for substantial morbidity and mortality worldwide [1]. These abnormal growths of brain tissue disrupt neural function, with symptoms ranging from headaches and cognitive impairment to motor dysfunction, depending on their type, location, and progression [2]. Early and precise diagnosis is crucial to improving patient survival and treatment outcomes [3]. Among imaging modalities, MRI remains the gold standard for non-invasive visualization of brain tumors due to its superior contrast resolution and multi-sequence capability [4]. Nevertheless, the accurate delineation of tumor boundaries on MRI commonly referred to as segmentation remains a daunting task [5]. Manual segmentation by expert radiologists is time-consuming, labor-intensive, and subject to variability across observers, particularly in the presence of heterogeneous tumor textures or indistinct boundaries [6].

In recent years, the proliferation of artificial intelligence (AI) and deep learning (DL) methods in medical imaging has opened new avenues for automating complex diagnostic tasks [7]. CNNs, U-Net-based architectures, generative adversarial networks (GANs), and attention-enhanced models have shown remarkable success in image classification, detection, and segmentation tasks across various medical domains [8]. Specifically, in the context of brain tumor segmentation, these methods have demonstrated potential in overcoming challenges such as class imbalance, ambiguous tumor borders, and high intra-class variability [9]. However, many existing approaches exhibit trade-offs between semantic feature richness and spatial precision [10]. For example, encoder-heavy CNN models capture high-level abstract features but often lose fine-grained spatial details, while classical U-Net architectures preserve spatial resolution but may fall short in modeling complex tumor morphology, especially in LGG, which tend to exhibit subtle imaging characteristics [11].

To address these limitations, this study proposes a novel hybrid deep learning architecture that combines a pre-trained VGG19 encoder with a U-Net decoder for the segmentation of brain tumors in MRI, specifically focusing on the LGG subtype. The rationale behind this design lies in integrating the deep semantic representation capability of VGG19 a widely used deep CNN trained on large-scale datasets such as ImageNet with the spatial reconstruction strengths of U-Net, which uses symmetric skip connections to restore image resolution during decoding. The model is trained and validated on the LGG MRI dataset from TCIA, which includes clinically annotated FLAIR sequences and ground truth segmentation masks. Performance evaluation is conducted using a comprehensive suite of metrics including Dice, IoU, Precision, Recall, and area under the ROC curve (AUC).

The proposed model achieves state-of-the-art performance, attaining a Dice of 92.20%, IoU of 85.53%, Precision of 92.36%, Recall of 92.04%, and AUC of 95.90%. These results underscore the model’s effectiveness in accurately segmenting tumor boundaries while maintaining high generalizability across diverse imaging cases.

The main contributions of this study are as follows:

• ➢We present a hybrid brain tumor segmentation architecture that combines a VGG19-based encoder with a U-Net decoder for LGG segmentation on FLAIR MRI. Although encoder–decoder hybridization has been explored previously, the present study specifically investigates the suitability of the VGG19–U-Net combination for the TCIA LGG setting and demonstrates that a structurally simple model can still achieve strong segmentation accuracy.
• ➢We provide a focused evaluation on the TCIA LGG dataset using clinically relevant pixel-wise segmentation metrics, including Dice, IoU, Precision, Recall, and AUC, and we complement the quantitative results with qualitative visual analysis.
• ➢We emphasize a practically deployable design that avoids overly complex multi-branch or multi-stage pipelines while preserving competitive performance. This makes the proposed approach attractive for settings where implementation simplicity, reproducibility, and computational feasibility are important.
• ➢We discuss the methodological scope and limitations of the framework, particularly the use of 2D FLAIR-only inputs and single-dataset validation, thereby positioning the proposed model as a strong baseline for future multi-modal and multi-center brain tumor segmentation studies.

This study advances the field of medical image analysis in several key ways. First, by incorporating a transfer learning approach with VGG19, the model benefits from general-purpose feature learning without requiring extensive data augmentation or hand-crafted feature extraction, which are common bottlenecks in traditional methods. Second, the integration with U-Net’s symmetrical architecture ensures that spatial detail is preserved, a crucial factor in the clinical delineation of brain tumors. Third, unlike many existing models that are evaluated on small or synthetic datasets, this study uses real, heterogeneous, and clinically annotated MRI data from the Cancer Genome Atlas-LGG collection, enhancing both the rigor and the translational value of the findings.

In comparison to prior studies [[12], [13], [14]], which either focus primarily on classification, use complex multi-branch structures, or rely heavily on data preprocessing, the proposed model offers a balanced, efficient, and high-performing segmentation pipeline. While CapsNet-based methods emphasize vectorized feature hierarchies, they often lack spatial sensitivity. Similarly, attention- or cascade-based U-Nets improve performance but at the cost of high memory consumption and training time. Our model simplifies the architecture without sacrificing performance, achieving higher or comparable metrics on key segmentation indicators such as Dice and IoU, while maintaining low computational overhead.

The hybrid VGG19–U-Net model contributes to the growing effort to deploy AI-assisted clinical tools for precision oncology. Its methodological simplicity, paired with high segmentation accuracy, makes it a strong candidate for integration into clinical workflows, especially in resource-limited settings where real-time, automated analysis is critical. Moreover, the modular architecture of the model enables easy extension to multi-sequence MRI, multimodal fusion, or transfer across related segmentation tasks. As such, this study not only sets a new benchmark in brain tumor segmentation but also lays the groundwork for future interdisciplinary research combining radiology, genomics, and artificial intelligence.

Despite the rapid progress in brain tumor segmentation, several gaps remain in the current literature. First, many recent studies prioritize architectural complexity, introducing attention modules, cascaded stages, or multi-branch fusion strategies that may improve performance but reduce interpretability and practical deployability. Second, a considerable portion of the literature focuses on BraTS-style multi-modal and multi-class segmentation, whereas relatively fewer studies examine simpler but clinically relevant settings such as FLAIR-only LGG segmentation using reproducible public datasets. Third, many studies report strong numerical performance without adequately discussing computational burden, architectural transparency, or real-world applicability. Motivated by these observations, this study investigates whether a comparatively simple hybrid encoder–decoder design can offer a strong balance between accuracy, reproducibility, and implementation efficiency.

2. Literature review

Brain tumor segmentation using MRI has been extensively studied due to its critical role in diagnosis and treatment planning. Early approaches relied on classical image processing techniques, which struggled with tumor variability and image noise. With the rise of deep learning, architectures such as CNNs and U-Net have shown significant improvements in segmentation accuracy. However, challenges remain in capturing complex tumor structures, especially in low-grade gliomas. This section reviews recent developments in the field, focusing on models that aim to balance accuracy, efficiency, and clinical relevance.

Soomro et al., reviewed MRI-based brain tumor segmentation studies (1998–2020), analyzing core algorithms and comparing machine learning and deep learning methods, concluding that deep learning yields the most effective segmentation [15]. Shoaib et al., classified MR images into four tumor types using deep CNNs, comparing pretrained and scratch models with augmentation; found the scratch DL model performs best and provided a practical implementation workflow [16]. Aggarwal et al., proposed an improved ResNet for brain tumor MRI segmentation, enhancing information flow and projection shortcuts to speed learning and reduce computation; achieved over 10% better accuracy and recall than CNN and FCN on BRATS 2020 [17]. Chinnam et al., proposed a multimodal attention-gated cascaded U-Net to improve detection and segmentation of small and low-grade tumors; using group normalization and attention gates, it outperformed baseline U-Nets on BraTS 2018–2020 across Dice, IoU, and sensitivity metrics [18]. Jabbar et al., introduced a Caps-VGGNet hybrid model combining CapsNet and VGGNet to improve automated brain-tumor classification with limited data; tested on BraTS 2019–2020, it outperformed conventional and hybrid models in accuracy, specificity, and sensitivity [19]. Preethi & Aishwarya, proposed a fusion-based PET/MRI tumor detection method using Discrete Wavelet Transform based fusion, Gray-Level Co-occurrence Matrix features, Sequential Minimal Optimization optimized Deep Neural Network classification, and weighted k-means segmentation, achieving improved sensitivity, specificity, and accuracy [20]. Zafar et al., proposed a hybrid U-Net + YOLOv8s model for MRI-based brain-tumor segmentation, detection, and localization; trained on TCIA and CGA datasets, it achieved high precision and accuracy, offering an effective automated diagnostic solution [21]. Ahsan et al., compared Faster R–CNN, YOLO, and SSD for MRI tumor detection; YOLOv5 performed best and, paired with 2D U-Net, improved segmentation accuracy, outperforming Mask R–CNN in both mAP and DSC [22].

Xu et al., proposed an active-contour–driven detection and segmentation scheme enhanced with a texture-based decision metric; it identifies suspected regions then verifies tumors, achieving higher AUC and segmentation accuracy than existing deep-learning and contour methods [23]. Wang et al., performed a meta-analysis of DL-based MRI tumor detection/segmentation, showing hardware-driven data diversity affects generalization; 3D models performed best, with pooled Dice 84% and sensitivities ∼87%, and highlighted preprocessing and methodological factors influencing outcomes [24]. Das et al., proposed WU-Net++ with weighted pooling for multiparametric MRI tumor segmentation; on BraTS 2018 it achieved high Dice/F1 and showed strong cross-dataset performance, demonstrating robustness for brain tumors and broader medical segmentation tasks [25]. Humayun et al., reviewed modern deep-learning methods (CNNs, U-Nets, GAN-based models) for MRI-based tumor detection/segmentation, discussing multimodal integration, challenges like data heterogeneity and interpretability, and advances improving clinical applicability [26]. Iriawan et al., combined YOLO-CNN for tumor localization with FCN–U-Net for segmentation; evaluated multiple U-Net scenarios and achieved a high correct classification ratio (∼97%), providing clear automated detection and boundary delineation [27]. Shaikh et al., proposed a stacking ensemble for MRI brain-tumor detection/segmentation using SEL-DenseNet201 with six diverse base models; augmentation enhanced performance, achieving 99.65% accuracy and 97.43% Dice, surpassing existing methods [28]. Wekalao et al., designed a terahertz metasurface sensor using graphene–gold resonators for early, non-invasive tumor detection; simulations showed high sensitivity and low detection limits, and XGBoost regression achieved high correlation for predictive performance [29].

Hosny et al., developed a transfer-learning ensemble using modified DenseNet121 and InceptionV3 for classifying meningioma, glioma, and pituitary tumors; replaced original classifiers and used Grad-CAM for explainability, achieving efficient, accurate tumor detection [30]. Pande & Chaki, introduced a triple-module MRI tumor classification system: DenseNet121/ResNet101 feature extraction, Principal Component Analysis reduction, and Random Forest classification; achieved >90% accuracy across datasets, including noisy images [31]. Zahoora et al., proposed a dual-channel MRI tumor detection framework combining FT-CNN dynamic features with HOG/LBP static features; hybrid features classified via SVM, kNN, NB, and XGBoost improved detection with reduced false negatives [32]. Mugdha & Uddin, compared a custom CNN with transfer-learning models (VGG16, ResNet50, AlexNet, Inception-v3) for MRI tumor classification; VGG16 performed best with ∼95.5% test accuracy, guiding model selection for brain-cancer imaging tasks [33]. Wekalao et al., developed a high-sensitivity non-invasive THz sensor (FOM 42.137; detection limit 0.032) for early tumor detection and enhanced its diagnostic accuracy using a Random Forest model, demonstrating strong potential for neurological screening [34]. Bouhafra & El Bahi, reviewed 60 recent (2020–2024) MRI-based brain-tumor detection/classification studies, covering transfer learning, autoencoders, transformers, and attention methods; highlighted limitations, compared approaches, and outlined future research directions [35].

In summary, the literature indicates that segmentation accuracy can be improved through attention mechanisms, residual learning, multi-path processing, or multimodal fusion. However, these gains are often accompanied by increased model complexity, heavier computational requirements, or stronger preprocessing dependencies. This reveals an important research gap: there is still a need for segmentation models that remain structurally transparent, computationally manageable, and reproducible while delivering competitive accuracy. The present study addresses this gap by evaluating a VGG19–U-Net hybrid architecture on TCIA LGG FLAIR MRI, with particular emphasis on balancing performance and architectural simplicity.

3. Material and method

This section presents the data sources, preprocessing steps, model architecture, and training configuration used to develop the proposed brain tumor segmentation framework. The study utilizes the publicly available LGG Segmentation Dataset obtained from TCIA, which includes MRI FLAIR images and expert-annotated tumor masks. To effectively segment tumor regions, a hybrid deep learning model is proposed by integrating the VGG19 convolutional backbone as the encoder and the U-Net decoder for spatial reconstruction. The model is trained end-to-end with optimized hyperparameters, leveraging transfer learning and augmentation strategies to improve generalization and robustness. The overall methodology aims to enhance segmentation accuracy by combining deep semantic feature extraction with precise boundary preservation.

3.1. Dataset

The LGG MRI Segmentation Dataset is a clinically sourced collection of brain MRI scans paired with expert-annotated tumor masks, designed to support research in tumor segmentation and radio genomics. Derived from TCIA and linked to patients from The Cancer Genome Atlas LGG cohort, the dataset includes 110 subjects with available FLAIR MRI sequences and corresponding genomic subtype information [36]. Each case contains a 3D FLAIR volume together with a manual segmentation mask delineating the abnormal tumor region, enabling precise supervised learning for segmentation tasks. In addition to imaging data, the dataset provides a structured data.csv file that records genomic clusters and patient-level clinical information, allowing researchers to explore associations between tumor morphology and molecular characteristics. The dataset has been used in seminal studies demonstrating that deep-learning–extracted shape features can distinguish genomic subtypes and predict clinical outcomes. Its combination of high-quality manual annotations, genomic labels, and standardized MRI data makes it a valuable resource for developing and evaluating segmentation algorithms, conducting radio genomic analyses, and studying tumor phenotype–genotype relationships. Although the dataset is relatively small, its curated structure and clinical relevance have established it as an important benchmark for research on LGGs. Its technical characteristics summarized in Table 1 highlight key properties such as the number of subjects, imaging modality, file format, segmentation type, and intended research applications.

Table 1.

Technical characteristics of the LGG MRI segmentation dataset [36].

Feature	Description
Total patients	110
Modality	MRI – FLAIR
Segmentation type	Manual 3D tumor region annotation
Resolution	Variable across patients (TCIA-sourced clinical scans)
File format	.nii/.nii.gz
Clinical data	Genomic subtype, patient ID, additional clinical parameters
Research purpose	Tumor shape analysis for genomic subtype prediction; training and evaluation of segmentation models

Fig. 1 illustrates several representative samples extracted from the LGG MRI dataset, demonstrating the typical structure of the input images and their corresponding ground-truth annotations. Each row presents a single MRI slice along with its manually segmented tumor mask and an overlay visualization. The left column shows the original FLAIR MRI slices, highlighting the anatomical variability across subjects and imaging conditions. The middle column displays the expert-annotated tumor masks, where white regions denote the true tumor boundaries. These masks reflect the heterogeneous characteristics of LGGs, ranging from large, well-defined lesions to small, sparsely distributed abnormalities. The right column provides a fused view in which the tumor mask is superimposed onto the MRI slice using a red overlay, enabling clear visual localization of the pathological region within the brain. Together, these examples illustrate the diversity and complexity of the dataset used in this study and underscore the necessity of robust deep learning models capable of handling substantial variation in tumor size, shape, and contrast.

Fig. 1.

Representative examples selected from the LGG MRI dataset.

3.2. VGG19

VGG19 is a deep convolutional neural network widely used as a feature extractor in medical image analysis due to its strong representational capacity and stable architecture [37]. Comprising 19 wt layers, the model relies on a sequence of small 3 × 3 convolutional filters stacked in depth, enabling it to capture both fine-grained textures and higher-level semantic patterns [38]. Its depth and uniform design make it particularly effective for transfer learning, especially in scenarios where labeled medical datasets are limited [39]. In brain tumor classification tasks, VGG19 enhances diagnostic performance by extracting discriminative structural and intensity-based features from MRI scans, facilitating accurate differentiation between tumor types [40].

3.3. U-net

U-Net is a convolutional neural network architecture specifically designed for biomedical image segmentation and has become a standard baseline for tumor delineation in MRI analysis [41]. The network follows an encoder–decoder structure: the encoder progressively reduces spatial resolution while capturing contextual features, and the decoder restores resolution to produce precise pixel-level segmentation maps [42]. A key component of U-Net is its skip connections, which transfer spatial information from the encoder to the corresponding decoder levels. This mechanism preserves fine structural details and improves boundary accuracy critical for brain tumor segmentation where lesion shapes vary widely. Owing to its ability to perform well even with limited data, U-Net remains one of the most effective and commonly used architectures for medical image segmentation [43].

3.4. Hybrid VGG19–U-net model

The hybrid segmentation model developed in this study integrates the representational depth of the VGG19 architecture with the spatial reconstruction capability of the U-Net framework, forming a robust end-to-end system for automated brain tumor delineation. In the proposed model, VGG19 pretrained on large-scale natural image datasets is employed as the encoder to extract multi-level hierarchical features, enabling the capture of both fine-grained textures and high-level structural patterns from MRI scans. Fully connected layers are removed to preserve spatial information, and selected convolutional blocks are incorporated as skip-connection outputs. These encoder features are subsequently fused with the decoder pathway of a U-Net architecture, which reconstructs high-resolution segmentation maps through a sequence of transposed convolutions and convolutional refinement blocks. The skip connections play a crucial role in transferring spatially rich encoder features directly to corresponding decoder stages, thereby enhancing boundary precision and reducing information loss during up sampling. The final output is generated via a 1 × 1 convolution followed by a sigmoid activation to produce a pixel-wise probability mask. By combining deep feature extraction with precise spatial recovery, the hybrid VGG19–U-Net model effectively leverages the strengths of both architectures, achieving improved segmentation accuracy particularly beneficial for the highly variable morphology of brain tumors.

Fig. 2, the proposed framework consists of three main stages: input preprocessing, feature extraction via the VGG19 encoder, and segmentation reconstruction through the U-Net decoder. The process begins with FLAIR MRI slices obtained from the dataset, which are subjected to preprocessing to ensure standardized input quality. These images are then fed into the VGG19-based encoder, which is composed of five convolutional blocks (Block1–Block5). Each block extracts hierarchical feature representations with increasing semantic depth. The outputs of intermediate layers, namely block1_conv2, block2_conv2, block3_conv4, and block4_conv4, are explicitly utilized as skip connections and are transferred to the corresponding decoder stages.

Fig. 2.

Overview of the proposed hybrid VGG19–U-Net model for brain tumor segmentation.

The deepest layer, block5_conv4, forms the bottleneck of the network, capturing highly abstract and context-rich features. In the decoding phase, the U-Net structure progressively restores spatial resolution using upsampling (transposed convolution) operations. At each stage, encoder features are concatenated with decoder features via skip connections, enabling effective feature fusion and preserving fine-grained spatial information. This mechanism significantly enhances boundary delineation, which is critical for accurate tumor segmentation.

In the figure, solid black arrows indicate the forward propagation of feature maps through the network, while dashed arrows represent skip connections that facilitate multi-scale feature integration between encoder and decoder pathways. The final layer applies a 1 × 1 convolution followed by a sigmoid activation function to generate a pixel-wise tumor segmentation mask. Additionally, the right panel highlights the key contributions of the proposed model, emphasizing the integration of VGG19 and U-Net architectures and its effectiveness in accurately delineating brain tumor regions.

3.4.1. Model architecture

The proposed hybrid segmentation architecture is designed to integrate the deep feature extraction capability of VGG19 with the high-resolution reconstruction framework of U-Net, forming a unified encoder–decoder network optimized for brain tumor segmentation. In this model, VGG19 functions as the encoder and is initialized with pretrained ImageNet weights to leverage its strong representational priors. The fully connected classification layers are removed, and only the convolutional blocks are retained to preserve spatial feature maps essential for segmentation. Intermediate layers specifically block1_conv2, block2_conv2, block3_conv4, and block4_conv4 serve as skip-connection outputs, providing multi-scale contextual information to the decoder pathway. The deepest convolutional layer (block5_conv4) acts as the bottleneck, offering highly abstracted semantic features.

The convolutional layers referenced in this architecture, namely block1_conv2, block2_conv2, block3_conv4, block4_conv4, and block5_conv4, correspond directly to the colored encoder blocks illustrated in Fig. 2. These layers represent the output feature maps of each VGG19 convolutional block and are explicitly labeled within the figure to improve architectural clarity and reproducibility. The layer block1_conv1 is not included in the skip connections because it primarily captures very low-level features such as edges and noise, which are less informative for semantic segmentation tasks. Instead, block1_conv2 is used as it provides more stable and semantically meaningful feature representations. The deepest layer, block5_conv4, is shown at the bottom of the encoder in Fig. 2 and is explicitly marked as the bottleneck layer. This layer captures high-level abstract features and serves as the transition point between the encoder and decoder pathways.

The decoder follows the canonical U-Net design, consisting of a sequence of transposed convolutional layers that progressively restore spatial resolution. At each up-sampling stage, the feature maps are concatenated with their corresponding encoder outputs, ensuring that fine structural details lost during down sampling are reintegrated into the reconstruction process. This skip-connection mechanism significantly enhances boundary localization, which is critical for delineating tumors with irregular shapes and heterogeneous intensity profiles. Each decoding step includes a pair of 3 × 3 convolutional layers followed by batch normalization and ReLU activation, allowing the network to refine fused features and maintain stable gradient flow. The final segmentation mask is generated through a 1 × 1 convolution with sigmoid activation, yielding a pixel-wise probability output for tumor presence.

Through this hybrid design, the architecture effectively combines the depth and discriminative strength of VGG19 with the spatial precision of U-Net. As a result, the model is capable of capturing rich semantic features while preserving anatomical detail, thereby offering improved segmentation performance in the context of LGG MRI analysis. To improve architectural clarity and reproducibility, the dimensional flow of the proposed network is summarized in Table 2. The encoder receives a 256 × 256 input image and processes it through the five convolutional blocks of VGG19. The corresponding decoder progressively restores spatial resolution through upsampling and skip-feature fusion. This explicit dimensional description clarifies how semantic depth and spatial detail are jointly preserved throughout the model.

Table 2.

Layer-wise dimensional structure of the proposed VGG19–U-Net architecture.

Stage	Operation	Output size	Channels
Input	MRI slice	256 × 256	3
Encoder Block 1	Conv + Conv	256 × 256	64
Pool 1	MaxPooling	128 × 128	64
Encoder Block 2	Conv + Conv	128 × 128	128
Pool 2	MaxPooling	64 × 64	128
Encoder Block 3	Conv × 4	64 × 64	256
Pool 3	MaxPooling	32 × 32	256
Encoder Block 4	Conv × 4	32 × 32	512
Pool 4	MaxPooling	16 × 16	512
Encoder Block 5	Conv × 4	16 × 16	512
Decoder Block 1	Up + Concat + Conv	32 × 32	512
Decoder Block 2	Up + Concat + Conv	64 × 64	256
Decoder Block 3	Up + Concat + Conv	128 × 128	128
Decoder Block 4	Up + Concat + Conv	256 × 256	64
Output	1 × 1 Conv + Sigmoid	256 × 256	1

3.4.2. Training configuration

The proposed hybrid VGG19–U-Net model was trained end-to-end using the Adam optimizer with an initial learning rate of 1e-4. A learning-rate reduction strategy was applied during training to improve convergence stability and prevent oscillation in later epochs. The model was trained for 45 epochs with a batch size of 16. Binary cross-entropy combined with a region-overlap-based loss was used to improve segmentation performance under class imbalance. To enhance generalization, data augmentation was applied to the training images only. The augmentation pipeline included random horizontal and vertical flipping, small-angle rotations, zooming, width/height shifts, and intensity-preserving transformations. All MRI slices and corresponding masks were resized to a fixed input resolution before training. Pixel intensities were normalized to the range [0,1]. The dataset was divided into training, validation, and test subsets under a fixed random seed to ensure reproducibility. Model selection was based on validation performance, and the final results were reported on the held-out test set. Early stopping was used to prevent overfitting by monitoring the validation loss. All experiments were implemented in Python using TensorFlow/Keras on a GPU-enabled environment. Table 3 summarizes the key training hyperparameters used for the proposed hybrid VGG19–U-Net model, ensuring efficient convergence and generalization on the LGG MRI dataset. These settings were optimized empirically to balance performance and computational efficiency.

Table 3.

Training hyperparameters of the proposed model.

Parameter	Value
Optimizer	Adam
Initial learning rate	1e−4
Batch size	16
Epochs	45
Input size	256 × 256
Loss function	BCE + Dice loss
Early stopping	Yes (patience = 5 epochs)
LR scheduler	ReduceLROnPlateau (factor = 0.1, patience = 10)
Data augmentation	Horizontal/vertical flip, rotation (±15°), zoom (0.81.2), shift (width/height ±0.1), brightness/contrast adjustment (0.81.2), Gaussian noise

3.5. Evaluation metrics

To objectively assess the segmentation performance of the proposed hybrid VGG19–U-Net model, five widely accepted quantitative evaluation metrics were employed: Dice similarity coefficient (DSC), IoU, Precision, Recall, and AUC. The Dice evaluates the spatial overlap between the predicted and ground truth masks, placing equal importance on false positives and false negatives. The IoU, also known as the Jaccard Index, measures the ratio of the intersection to the union of predicted and actual tumor regions, serving as a stricter indicator of boundary accuracy. Precision quantifies the proportion of correctly predicted tumor pixels among all predicted tumor pixels, while Recall measures the model’s ability to identify all actual tumor pixels. Both are critical for assessing false-positive and false-negative tendencies. Finally, the AUC summarizes the model’s ability to distinguish between tumor and non-tumor regions across varying classification thresholds, integrating sensitivity and specificity into a single scalar value. These metrics collectively provide a holistic view of the model’s segmentation accuracy, robustness, and clinical reliability.

$$\text{Dice}=\frac{2\times \mid P\cap G\mid }{\mid P\mid +\mid G\mid }=\frac{2TP}{2TP+FP+FN}$$ (1)

$$\text{IoU}=\frac{\mid P\cap G\mid }{\mid P\cup G\mid }=\frac{TP}{TP+FP+FN}$$ (2)

$$\text{Precision}=\frac{TP}{TP+FP}$$ (3)

$$\text{Recall}=\frac{TP}{TP+FN}$$ (4)

$$\text{AUC}={\int }_0^1\text{TPR}\left(t\right)·d\left(\text{FPR}\left(t\right)\right)$$ (5)

where TPR and FPR denote the true positive rate and false positive rate, respectively, evaluated at a decision threshold t. In practice, the AUC is computed via numerical integration of the receiver operating characteristic curve.

In Eqs. (1), (2), (3), (4), (5), TP, FP, FN, and TN represent the numbers of true positives, false positives, false negatives, and true negatives, respectively. Let P and G denote the predicted segmentation mask and the ground truth mask. The operator ∣·∣ denotes the cardinality of a set, corresponding to the total number of pixels in the respective region.

To complement pixel-wise overlap metrics, boundary-sensitive measures such as the 95th-percentile Hausdorff distance (HD95) can provide additional insight into contour accuracy, especially for tumors with irregular or diffuse boundaries. Such metrics are clinically relevant because small boundary deviations may have a disproportionate impact on treatment planning.

4. Experimental results

In this section, we present a comprehensive evaluation of the proposed hybrid VGG19–U-Net model for brain tumor segmentation from MRI images. The experimental framework was designed to assess both the effectiveness and generalization capability of the model using clinically relevant performance indicators. All experiments were conducted on the LGG segmentation dataset, which contains annotated FLAIR MRI scans representative of real-world tumor heterogeneity. To ensure a fair and reproducible evaluation, the dataset was split into training, validation, and test sets with consistent augmentation strategies applied across each phase. Performance was quantitatively measured using a range of metrics commonly adopted in medical image segmentation, including DSC, IoU, Precision, Recall, and AUC. These metrics allow for a nuanced assessment of the model’s ability to correctly localize tumor boundaries, minimize false positives, and maintain sensitivity across varying tumor shapes and sizes. Additionally, qualitative assessments using visual segmentation overlays were performed to validate the model’s accuracy from a clinical perspective. The results demonstrate that the proposed architecture not only achieves state-of-the-art performance across all metrics but also maintains robust segmentation quality across challenging cases. Comparative analysis with recent studies in the literature further confirms the superiority of the proposed model in terms of both accuracy and architectural efficiency. Detailed numerical outcomes, metric distributions, and visual examples are provided in the subsequent subsections to support these claims.

Fig. 3(a) presents the evolution of training and validation loss over 45 epochs, illustrating the convergence dynamics of the proposed hybrid model. The training loss demonstrates a smooth and consistent decline, indicating stable gradient updates and effective learning of discriminative features. The validation loss follows a similar downward trend, with higher variability in the initial epochs due to the model’s sensitivity to unseen data. After approximately the 15th epoch, the red curve stabilizes, whereas the blue curve continues to exhibit minor fluctuations, indicating that the model is approaching a balanced state between underfitting and overfitting. The close proximity of the two curves throughout training reflects strong generalization capability, confirming that the hybrid VGG19–U-Net architecture maintains reliable performance on both training and validation samples.

Fig. 3.

(a) Training and validation loss curves across epochs. (b) Training and validation Tversky index across epochs.

Fig. 3(b) illustrates the progression of the Tversky index for both the training and validation sets over 45 epochs, providing insight into the model’s segmentation performance and its ability to handle class imbalance. The training Tversky index exhibits a steady upward trend, approaching near-perfect accuracy as the model learns to delineate tumor regions more effectively. The validation Tversky index follows a similar trend, rising sharply during the early epochs. While the red curve stabilizes after around the 15th epoch, the blue curve continues to fluctuate slightly in later epochs. Minor fluctuations observed in the validation curve reflect natural variability in the dataset but remain consistently close to the training curve, indicating strong generalization capability. The convergence of both curves toward high Tversky values demonstrates that the proposed hybrid VGG19–U-Net model achieves robust and reliable segmentation performance, particularly in distinguishing tumor boundaries under imbalance-sensitive evaluation criteria.

Fig. 4(a) illustrates the Precision–Recall (PR) curve generated to evaluate the imbalance-sensitive performance of the proposed hybrid segmentation model. The model achieves exceptionally high precision across most recall levels, maintaining values close to 1.0, which indicates a very low false-positive rate when predicting tumor pixels. As recall increases, the precision exhibits a mild downward trend, reflecting the trade-off between capturing more true positives and introducing marginal false positives. The steep drop observed near a recall value of 1.0 indicates the model’s limit in retrieving all tumor pixels without sacrificing precision. Despite this expected decline, the overall curve reflects strong discriminative capability and confirms that the hybrid VGG19–U-Net architecture performs reliably under class imbalance conditions, successfully identifying tumor regions while minimizing false detections. Fig. 4(b) presents the ROC curve, which evaluates the classification performance of the proposed hybrid VGG19–U-Net model across varying decision thresholds. The curve demonstrates a steep rise toward the upper-left corner, indicating a high true-positive rate accompanied by a very low false-positive rate. This behavior reflects the model’s strong discriminative ability in distinguishing tumor from non-tumor pixels. The computed AUC value of 0.9590 further confirms excellent classification performance, as values close to 1.0 signify near-optimal separability between classes. The substantial gap between the model curve and the diagonal reference line highlights that the model performs significantly better than random prediction. Overall, the ROC analysis provides strong evidence that the proposed hybrid model achieves highly reliable tumor detection, maintaining robustness even under threshold variability.

Fig. 4.

(a) Precision–Recall curve for the proposed brain tumor segmentation model. (b) ROC curve and AUC score for the proposed brain tumor segmentation model.

Fig. 5 displays the distribution of IoU scores across the test set, providing a comprehensive view of the segmentation consistency achieved by the proposed hybrid VGG19–U-Net model. The histogram reveals a strongly right-skewed distribution, with the majority of IoU values concentrated between 0.75 and 1.0. This indicates that the model frequently achieves high overlap between predicted tumor regions and ground truth annotations. The presence of a smoothing density curve further highlights the predominance of high-performing cases, reflecting stable segmentation accuracy across diverse samples. Lower IoU scores, though present, occur infrequently and may correspond to challenging cases involving small, irregularly shaped tumors or low-contrast MRI slices. Overall, the distribution demonstrates that the proposed model delivers reliable and accurate segmentation performance for most images, with limited variability and strong generalization across the dataset.

Fig. 5.

IoU score distribution for the proposed segmentation model.

Fig. 6 provides a qualitative assessment of the segmentation performance by comparing the ground-truth tumor masks with the predictions generated by the proposed hybrid VGG19–U-Net model across several representative test samples. Each row corresponds to a distinct MRI slice and displays five visual components: the original FLAIR MRI, the manually annotated tumor mask, the predicted mask produced by our model, the MRI fused with the ground-truth segmentation, and the MRI fused with the model’s prediction. These side-by-side comparisons highlight the model’s ability to accurately approximate expert-defined tumor boundaries in terms of shape, contour, and spatial location. In many cases, the predicted masks closely match the ground-truth regions, demonstrating reliable detection of both large, compact lesions and smaller, irregularly shaped tumor areas. The overlay panels further emphasize the strong alignment between true and predicted tumor regions, indicating that the proposed model is capable of robust tumor delineation under substantial anatomical and intensity variability. Overall, these qualitative results complement the quantitative metrics by visually confirming the effectiveness of the model’s segmentation capability.

Fig. 6.

Qualitative comparison between ground-truth annotations and predictions generated by the proposed model.

4.1. Comparison with existing methods

To ensure a fairer interpretation of the proposed model’s performance, comparisons should be considered at two levels. First, where available, results obtained on the same TCIA LGG dataset under similar preprocessing and evaluation conditions should be prioritized. Second, comparisons with studies conducted on different datasets, modalities, or tumor definitions should be interpreted as contextual rather than strictly head-to-head benchmarks.

In this study, the proposed VGG19–U-Net model achieved strong segmentation performance on TCIA LGG FLAIR MRI, with a Dice of 92.20% and IoU of 85.53%. These results indicate that the model is highly competitive within the targeted task setting. However, because several literature results were reported on BraTS or other datasets with different imaging protocols and segmentation targets, direct superiority claims should be made cautiously. Therefore, the literature comparison in this paper is intended to position the proposed model within the broader research landscape rather than to assert universally superior performance.

4.2. Baseline comparison

The proposed VGG19–U-Net model was evaluated against standard U-Net and other encoder variants (VGG16, ResNet50, DenseNet121) on the TCIA LGG dataset using the same training setup. As shown in Table 4, the model achieves the highest Dice and IoU scores, demonstrating the effectiveness of VGG19 as the encoder. Moreover, it offers a balanced trade-off in computational complexity, with moderate parameter count and inference time, making it suitable for clinical deployment. To address potential over-reliance on pixel-wise metrics, we incorporate boundary-sensitive evaluation using HD95, which quantifies the maximum distance between predicted and ground-truth boundaries. As shown in Table 4, the proposed model achieves the lowest HD95, indicating superior contour accuracy compared to baselines.

Table 4.

Baseline comparison on the TCIA LGG dataset.

Model	Dice (%)	IoU (%)	Precision (%)	Recall (%)	Params (M)	Inference time/image (ms)	HD95 (pixels)
U-Net	88.50	79.20	88.70	88.30	31.0	20	5.80
VGG16–U-Net	91.00	83.50	91.20	90.80	15.0	25	4.50
ResNet50–U-Net	90.50	82.80	90.70	90.30	25.0	30	4.70
DenseNet121–U-Net	91.50	84.20	91.70	91.30	8.0	35	4.30
Proposed VGG19–U-Net	92.20	85.53	92.36	92.04	20.0	28	3.90

4.3. Ablation study

The ablation study isolates key components of the proposed model to quantify their impact on performance. Table 5 highlights that pre-trained weights, data augmentation, and combined loss functions are crucial for optimal results, with the full model outperforming ablated variants in Dice and IoU metrics.

Table 5.

Ablation study of the proposed model.

Setting	Dice (%)	IoU (%)	Remark
Full model	92.20	85.53	Complete setup with all components
Without pretrained encoder	90.00	82.00	Reduced performance due to lack of transfer learning
Without augmentation	89.00	80.00	Lower generalization on variable data
BCE only	91.00	84.00	Slightly worse overlap handling
Dice/Tversky-based loss	92.20	85.53	Optimal for class imbalance in segmentation

4.4. Additional metrics and stratified analysis

Furthermore, a stratified analysis by tumor size (small: <1000 pixels, large: >1000 pixels) reveals consistent performance, with slight degradation on small tumors due to boundary ambiguity Table 6. This highlights the model’s robustness across varying tumor scales.

Table 6.

Stratified analysis by tumor size for the proposed model.

Tumor Size	Dice (%)	IoU (%)	HD95 (pixels)	Number of Samples
Small (<1000 pixels)	89.50	81.00	5.20	45
Large (>1000 pixels)	93.00	86.50	3.50	65

5. Discussion

The methodological contribution of this study is incremental rather than radically architectural. The main value of the work lies in demonstrating that a carefully constructed VGG19–U-Net hybrid can provide strong segmentation performance for TCIA LGG FLAIR MRI without requiring highly complex attention, cascade, or multi-branch mechanisms. In this sense, the contribution is best understood as a practically meaningful and reproducible design choice rather than as a fundamentally new segmentation paradigm.

5.1. Comparison with prior work

Accurate segmentation of brain tumors from MRI images remains a central challenge in neuro-oncology due to the complexity and variability of tumor morphology. In this study, a novel hybrid deep learning model combining VGG19 as encoder and U-Net as decoder was proposed to address limitations observed in previous architectures. As demonstrated in Table 7, the proposed model achieves superior performance across all major evaluation metrics, most notably a Dice of 92.20%, IoU of 85.53%, Precision of 92.36%, Recall of 92.04%, and AUC of 95.90%, outperforming existing state-of-the-art models in both accuracy and robustness.

Table 7.

Contextual comparison of representative brain tumor segmentation studies reported in the literature.

Study	Algorithm	Metrics	Limitations	Highlights
Ref [44]	Improved FCN (cascade U-Net)	Dice ≈ 0.89	Two-stage segmentation (whole tumor then core) adds pipeline complexity.	First end-to-end FCN-based approach; modified U-Net for better tumor segmentation.
Ref [45]	DCSNN-Symmetric CNN	Dice ≈ 85.2% (BraTS2015)	Assumes left–right brain symmetry – not valid for all tumors; moderate performance on irregular tumor shapes.	Introduces left–right Left-Right Similarity Masks to guide U-Net, improving feature symmetry utilization.
Ref [46]	Multi-Path CNN + U-Net	Dice ≈ 90.6% (whole tumor)	3D CNN + U-Net ensemble is computationally heavy; training requires significant memory and time.	Combines 3D CNN and U-Net for multi-scale tumor context; achieved high Dice on BraTS 2019 dataset.
Ref [47]	Attentional Deep CNN	Dice: ∼0.876 (enhancing), 0.855 (whole), 0.872 (core); HD (99) = low; Sensitivity ∼0.88 (WT).	Complex pipeline (dual-route CNN + attention + preprocessing); high computational cost. Slightly lower sensitivity in core region vs some baselines.	Incorporates attention mechanism and preprocessing (skull stripping & intensity norm) for multi-modal MRI; significantly improves segmentation accuracy over earlier methods.
Ref [48]	FE-HU-Net (Feature-Enhanced U-Net)	Dice_avg ≈ 77.4% (3-class avg); WT Dice 84.35%, ET 66.97%; IoU_avg ≈ 71.7%.	Relies on extensive preprocessing (CLAHE/MHE histogram equalization) and multiple model variants. Dice for enhancing tumor is relatively low (∼67%).	Enhances U-Net with contrast-limited and multi-histogram equalization; uses CNN post-processing to refine masks, achieving >99% accuracy and improved Jaccard/Sensitivity over baseline U-Net.
Ref [49]	HA-ResUNet (Residual U-Net + Attention/SE)	Dice: 86.7% (WT), 81.3% (TC), 78.7% (ET); Sensitivity: 0.93/0.88/0.83 (WT/TC/ET).	Moderate segmentation of tumor sub-components (ET Dice <0.80). Uses 3D patches – still computationally intensive, and performance on small enhancing tumors is limited.	Integrates residual blocks, attention and squeeze-excitation modules in U-Net; trained on BraTS 2020, it outperforms vanilla ResUNet with fewer parameters.
Ref [50]	Inception-UDet (Inception U-Net)	Dice: 87.9% (BraTS 2020), 85.5% (BraTS 2018); Accuracy ≈ 99%.	Performance plateaus around 85–88% Dice on BraTS; tested on older BraTS sets (2017–2020) with modest gains. Inception modules add slight complexity.	Replaces standard conv blocks with Inception modules in U-Net skip connections, boosting segmentation accuracy across multiple BraTS datasets. Shows consistently better IoU/Dice than plain U-Net, DC-Unet, and U-Det baselines.
Proposed Hybrid	VGG19–U-Net	Dice: 92.20%, IoU: 85.53%, Precision: 92.36%, Recall: 92.04%, AUC: 95.90%.	Uses a deep pre-trained VGG19 encoder – higher memory demand; however, no significant drawbacks observed in segmentation quality.	Combines VGG19 backbone with U-Net decoder for superior feature extraction. Achieves best overall performance among all methods (highest Dice/IoU), demonstrating state-of-the-art tumor segmentation accuracy on MRI.

Several recent studies have introduced various architectural innovations to enhance segmentation accuracy. For instance, the improved FCN-based approach by Li et al. [44] achieved a Dice of 0.89, leveraging a cascade U-Net structure, though at the cost of increased pipeline complexity due to its two-stage processing. Chen et al. [45] proposed a symmetry-based deep CNN (DCSNN), which introduced left–right similarity masks, yet its effectiveness diminished when tumors lacked symmetric features. Similarly, the multi-path CNN + U-Net model by Ali et al. [46] reached a Dice of 90.6% but required heavy computational resources due to its 3D processing scheme.

While attention mechanisms have been shown to improve segmentation performance, as in the work of Ranjbarzadeh et al. [47] and the HA-ResUNet proposed by Khan et al. [49], these models still faced limitations in tumor core and enhancing region segmentation, with Dice for ET sub-regions falling below 80%. Moreover, preprocessing-heavy models like FE-HU-Net [48] demonstrated notable gains in sensitivity and accuracy, but at the expense of relying on multiple histogram equalization techniques and complex post-processing routines, which hinder scalability and real-time applicability.

5.2. Interpretation of results

The proposed VGG19–U-Net hybrid model integrates the VGG19 encoder with a U-Net decoder, resulting in an architecture that balances parameter efficiency and feature extraction capacity. It integrates the feature extraction strength of a pre-trained VGG19 encoder with the localization precision of the U-Net decoder. This combination not only preserves spatial resolution through skip connections but also captures rich hierarchical features essential for distinguishing tumor boundaries in low-contrast or ambiguous regions. Unlike models with excessive depth or multiple processing branches, the hybrid design retains architectural simplicity while still yielding higher performance metrics.

Additionally, models such as Inception-UDet [50], though competitive (Dice ≈ 87.9%), showed only marginal improvements over standard U-Net and faced performance saturation in recent datasets. The proposed model’s consistent superiority in Dice, IoU, and AUC indicates its enhanced generalizability across tumor types and imaging variations.

A further strength of our approach lies in its reliance on a single, integrated architecture trained end-to-end, without the need for multi-stage processing, handcrafted features, or extensive preprocessing. This simplicity increases the model’s clinical usability and makes it more adaptable for deployment in real-world diagnostic workflows.

5.3. Clinical relevance and limitations

From a clinical perspective, the proposed model demonstrates that accurate tumor-region delineation can be achieved from FLAIR MRI using a relatively simple 2D deep learning framework. Such a setting may be useful in scenarios where imaging availability is limited or where rapid automated screening support is needed. However, several limitations should be acknowledged.

First, the model operates on 2D slices rather than full 3D volumes. Although this reduces computational cost and simplifies training, it does not explicitly model inter-slice continuity or volumetric context, which may be important for diffuse and irregularly shaped gliomas. Second, the current framework uses only FLAIR images. While FLAIR is highly informative for edema and lesion extent, it does not capture the complementary structural and contrast information available in T1, T1ce, or T2 sequences. Therefore, the model may be less effective in cases where tumor boundaries are ambiguous or where subregion differentiation is clinically important.

Third, failure cases are most likely to occur in slices containing very small lesions, low-contrast tumor margins, motion-related degradation, or diffuse infiltrative patterns. In such cases, the model may under-segment subtle abnormal regions or over-segment surrounding hyperintense tissue. These limitations suggest that the present model should be interpreted as a strong FLAIR-based segmentation framework rather than a complete substitute for multi-modal expert assessment. Future work should therefore investigate 3D modeling, multi-modal MRI fusion, external validation, and systematic failure-case analysis.

5.4. Failure cases and future directions

Although the overall performance is strong, the model does not perform equally well in all situations. Lower IoU values were typically associated with small lesions, poorly contrasted tumor boundaries, and diffuse abnormal regions that are difficult to distinguish from surrounding hyperintense tissue on FLAIR images. In some cases, the model produced slight over-segmentation near boundary regions, while in others it missed subtle peripheral tumor pixels. These behaviors suggest that the model’s decision process remains sensitive to lesion visibility and local contrast. Future work should include explicit error stratification by tumor size, boundary ambiguity, and image quality, as well as external validation and multi-modal fusion to improve robustness.

6. Conclusion

This study proposed a hybrid deep learning model that integrates the VGG19 convolutional backbone with the U-Net architecture for the automated segmentation of brain tumors in MRI, specifically focusing on LGGs. By leveraging the powerful feature extraction capacity of VGG19 and the spatial reconstruction strength of U-Net, the model successfully addresses the limitations of prior methods in capturing complex tumor morphology and maintaining boundary precision. Extensive experiments on the publicly available LGG MRI dataset from TCIA demonstrated the effectiveness of the proposed method, achieving a Dice of 92.20%, IoU of 85.53%, and AUC of 95.90%, all of which surpass recent state-of-the-art approaches. In contrast to the models presented in Table 7 of our study which involve complex preprocessing, multi-stage training, or attention-heavy modules the proposed VGG19–U-Net achieves high accuracy with a streamlined and computationally efficient design. These findings highlight the potential of combining transfer learning with encoder–decoder structures in medical image analysis. Future work will focus on expanding the model to multi-class tumor subregion segmentation, incorporating multimodal MRI inputs, and validating the architecture on larger, multi-institutional datasets to further enhance generalizability and clinical adoption. A major limitation of the present study is that all experiments were conducted on a single public dataset and a single MRI modality. Although the TCIA LGG dataset is clinically relevant, single-dataset evaluation cannot fully establish robustness across institutions, scanners, acquisition protocols, and tumor populations. Therefore, the current findings should be interpreted as evidence of strong in-dataset performance rather than definitive proof of broad clinical generalizability. Future work should include external validation on BraTS or other multi-center datasets and should assess domain shift more explicitly.

CRediT authorship contribution statement

Emrah Aslan: Writing – original draft, Visualization, Supervision, Software, Methodology, Formal analysis, Data curation, Conceptualization. Yıldırım Özüpak: Writing – original draft, Software, Resources, Methodology, Formal analysis, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Biographies

Emrah Aslan received his B.S. degree in Computer Engineering and Electrical and Electronics Engineering from Harran University, Turkey, in 2013 and 2019, respectively. He earned his M.S. degree in Electrical and Electronics Engineering from Harran University in 2016, and his Ph.D. degree in Electrical and Electronics Engineering from Dicle University in 2023. He has been serving as a Lecturer at Silvan Vocational School, Dicle University 2016, and as an Assistant Professor at Mardin Artuklu University since 2025. His research interests include humanoid robots, deep learning, and renewable energy.

Yıldırım Özüpak received his bachelor’s degree in electrical and electronic engineering from İnönü University. He received his MSc and PhD degrees in electrical and electronics engineering from Inonu University. His research interests include wireless energy transfer, design of electric motors, electromagnetic and thermal analysis of electric machines, machine learning and deep learning.