Neuroscience · Single study

Deep learning model classifies brain tumors from MRI scans with 98% accuracy

• Researchers developed an artificial intelligence system that identifies brain tumors in MRI images by combining two techniques: federated learning, which trains the model across multiple locations without sharing private medical data, and transfer learning, which adapts knowledge from other imaging tasks.
• The system was trained on thousands of brain MRI scans and achieved 98% overall accuracy, correctly identifying gliomas, meningiomas, pituitary tumors, and normal brains in test cases.
• This approach could help radiologists diagnose brain tumors faster and more reliably while keeping patient data private during the training process.

BMC Medical Imaging · 2024 — https://doi.org/10.1186/s12880-024-01261-0

Integrated approach of federated learning with transfer learning for classification and diagnosis of brain tumor

Introduction

Brain tumors pose intricate challenges due to their location in the delicate structure of the human brain. These abnormal masses of cells, which can be either benign or malignant, present a wide spectrum of complexities that extend beyond their classification. Understanding these complexities is crucial in comprehending the gravity of the condition and the intricacies of treatment.

Benign tumors, though noncancerous, can still cause significant issues depending on their location and size. They may exert pressure on the brain, leading to symptoms such as headaches, seizures, or neurological deficits. However, these tumors typically have well-defined borders and tend to grow slower than malignant tumors. Surgical removal might offer a curative option for these tumors, although their location within critical brain regions might limit the feasibility of complete resection without causing damage to essential brain structures.

In contrast, malignant brain tumors, also known as brain cancer, exhibit more aggressive behavior. They grow rapidly and infiltrate surrounding healthy brain tissue, making complete surgical removal challenging. The most common malignant primary brain tumor in adults is glioblastoma multiforme, notorious for its aggressive nature and resistance to treatment. Its diffuse nature within the brain makes it challenging to eradicate entirely, leading to a high recurrence rate despite aggressive treatment approaches involving surgery, radiation, and chemotherapy.

The diversity of brain tumors further complicates treatment strategies. There are distinct types of tumors, such as gliomas, meningiomas, pituitary adenomas, and medulloblastomas, each with their unique characteristics and challenges. For instance, some tumors originate from the brain tissue itself, while others may develop from surrounding structures or metastasize from cancers elsewhere in the body. This diversity demands tailored approaches for accurate diagnosis, prognosis, and treatment planning.

The skull serves as an unyielding shield, guarding the brain against external forces. However, this rigid structure becomes a hindrance when faced with internal growth, whether benign or malignant. Brain tumors, regardless of their nature, can pose severe challenges due to the limited space within the skull. Their presence often leads to heightened intracranial pressure, a condition that can culminate in brain damage or life-threatening situations.

The World Health Organization (WHO) adopts a systematic classification system for brain tumors, aiming to categorize them based on their type, level of malignancy, and grade. This categorization is pivotal in guiding the treatment approach and understanding the prognosis associated with each tumor type.

The skull’s rigidity means that any growth within this confined space can trigger a cascade of issues. Even benign tumors, while not cancerous, can exert substantial pressure on the brain as they grow. Their expansion within the limited confines of the skull can lead to a rise in intracranial pressure, which, in turn, might cause symptoms ranging from persistent headaches to nausea, vomiting, seizures, and even neurological deficits.

Malignant tumors, on the other hand, present a graver concern. Their aggressive nature, characterized by rapid growth and invasive tendencies, exacerbates the challenges posed by limited intracranial space. As these tumors progress, they infiltrate and displace healthy brain tissue, amplifying the elevation of intracranial pressure. This situation can quickly escalate, causing severe neurological impairment or life-threatening consequences if not managed promptly and effectively.

The WHO classification system for brain tumors is a vital tool in understanding the diverse landscape of these conditions. It categorizes tumors into several types based on their cellular origin, characteristics, and behavior. Moreover, it differentiates between grades, reflecting the level of malignancy and the tumor’s aggressiveness.

For instance, gliomas, a type of tumor originating from glial cells, encompass a spectrum ranging from low-grade (such as astrocytoma’s and oligodendrogliomas) to high-grade tumors like glioblastoma multiforme (GBM), known for their aggressive behavior. Meningiomas, arising from the meninges, are typically categorized as benign tumors, but depending on their location and growth pattern, they too can exert pressure on the brain and cause significant issues.

The WHO grading system further stratifies tumors based on their histopathological features, aiding clinicians in determining the prognosis and guiding treatment decisions. Grade I and II tumors are considered low-grade, often growing slowly, and possessing more defined borders, while Grade III and IV tumors represent high-grade malignancies, exhibiting rapid growth and infiltrative tendencies.

Magnetic Resonance Imaging (MRI) stands as a cornerstone in diagnosing brain tumors due to its ability to offer highly detailed images of the brain’s anatomy. However, interpreting these images accurately to diagnose and classify brain tumors poses a complex challenge. Traditionally, this task has relied on the expertise of radiologists, yet this manual interpretation is time-consuming, subjective, and susceptible to human error, especially in intricate cases or when managed by less experienced personnel.

The emergence of machine learning, particularly deep learning techniques, has revolutionized medical image analysis, presenting novel prospects for brain tumor diagnosis. Convolutional Neural Networks (CNNs), a type of deep learning algorithm, have exhibited remarkable potential in precisely categorizing images, including those from medical imaging. Their adeptness in learning intricate patterns and features from vast datasets renders them suitable for tasks like brain tumor classification. The types of Brain Tumors are discussed in Table.

Table 1. Types of brain tumors

The contemporary methodologies for brain tumor classification harness deep learning by training CNN models on extensive datasets comprising MRI images. These models are trained to discern and identify patterns and features associated with several types of brain tumors. Despite notable advancements, challenges persist in terms of data privacy, model generalization, and the demand for substantial, annotated datasets.

MRI’s unparalleled ability to produce high-resolution images of the brain enables detailed visualization of tumors, providing crucial information for diagnosis and treatment planning. However, the process of analyzing these images manually relies heavily on radiologists’ expertise, leading to subjectivity and potential errors. Moreover, interpreting complex MRI images to differentiate between various tumor types demands a profound understanding of subtle nuances that might not always be evident to the human eye.

The integration of deep learning techniques, especially CNNs, has shown immense promise in revolutionizing the interpretation of MRI images for brain tumor diagnosis. These algorithms can autonomously learn intricate patterns and features within images, potentially augmenting the accuracy and efficiency of tumor classification.

CNNs function by utilizing multiple layers to detect hierarchical patterns within images. They learn from large volumes of labeled data, gradually enhancing their ability to recognize specific features associated with diverse types of brain tumors. This learning process involves the extraction of features at various levels of abstraction, enabling the network to discern subtle variations indicative of distinct tumor characteristics.

Despite the considerable progress made with CNNs, challenges persist within this domain. Data privacy remains a concern due to the sensitive nature of medical imaging data. Annotated datasets, crucial for training deep learning models, are often limited in size and accessibility due to privacy regulations and the labor-intensive nature of labeling medical images.

Furthermore, ensuring the generalizability of these models beyond the datasets they were trained on remains a significant challenge. Models trained on specific datasets might encounter difficulties when applied to new, unseen data or when faced with variations in imaging techniques or equipment.

Efforts to address these challenges include the development of privacy-preserving techniques that enable model training without compromising patient data confidentiality. Transfer learning, a method where pre-trained models are fine-tuned with smaller datasets, offers a potential solution for mitigating the need for vast amounts of annotated data. Additionally, collaborations between healthcare institutions for data sharing and the creation of standardized datasets could facilitate model training and validation across diverse populations [].

The integration of deep learning in brain tumor classification using MRI images signifies a promising avenue in improving diagnostic accuracy and efficiency. As technology advances and methodologies evolve, the synergy between machine learning and medical imaging is poised to enhance our ability to detect, classify, and manage brain tumors, potentially transforming patient care and outcomes. However, addressing challenges related to data privacy, model generalization, and dataset availability will be crucial in realizing the full potential of these advancements in clinical practice [].

To address these challenges, we propose a novel federated learning-based deep learning model for automated and accurate brain tumor classification. This innovative approach not only emphasizes the use of a modified VGG16 architecture optimized for brain MRI images but also highlights the significance of federated learning and transfer learning in the medical imaging domain. Federated learning enables decentralized model training across multiple clients without compromising data privacy, addressing the critical need for confidentiality in medical data handling. Additionally, transfer learning leverages a pre-trained CNN, enhancing the model’s ability to classify brain tumors accurately by leveraging knowledge gained from vast and diverse datasets.The Contribution of the Research Paper are:

The subsequent sections of the research paper encompass vital facets crucial for a comprehensive study. The “Related Work” segment intricately surveys existing technologies, offering a detailed overview of prevailing methodologies. Following this, the “Materials and Methods” section elaborates on the dataset used, the CNN model architecture, and the innovative federated learning approach. The “Results” segment showcases empirical findings, spotlighting the model’s performance via metrics like accuracy, precision, recall, and F1-scores. Subsequently, the “Discussion” section conducts a thorough analysis, comparing outcomes with established methods, exploring implications, and addressing study limitations. The “Conclusion” succinctly summarizes key findings’ potential impacts on medical diagnostics and delineates avenues for future research. Lastly, the “References” compile all referenced scientific literature and data sources, ensuring academic integrity, and acknowledging scholarly contributions.

Related work

The field of medical imaging, particularly the classification and diagnosis of brain tumors using MRI images, has seen significant advancements with the integration of machine learning and deep learning techniques. This section reviews related work in this domain, highlighting key methodologies, findings, and how they relate to our current research.

Our research builds upon and extends these developments by proposing a federated learning-based deep learning model, utilizing a modified VGG16 architecture for the classification of brain tumors from MRI images. This approach not only addresses the limitations associated with traditional techniques, machine learning, and deep learning methods but also leverages the strengths of federated learning to offer a novel solution that prioritizes precision, efficiency, and data privacy. By comparing with the existing methodologies outlined in Table, our study contributes a unique perspective to the ongoing dialogue in this rapidly evolving field, highlighting the potential of federated learning to overcome some of the most pressing challenges in healthcare applications.

Table 2. Existing methodologies

By comparing with these related studies, our research contributes to the ongoing dialogue in this rapidly evolving field, offering a novel approach that balances the need for precision, efficiency, and data security in healthcare applications.

Methodology

This research employs an advanced machine learning approach, combining Convolutional Neural Networks (CNNs) with a federated learning framework, to classify brain tumors using Magnetic Resonance Imaging (MRI) data. The methodology encompasses several key components: dataset preparation, model architecture design, federated learning implementation, training procedures, and evaluation metrics. The architecture of the proposed model has been given in Fig..

Fig. 1. Proposed model

A. Dataset description and preparation

The research utilizes a comprehensive dataset comprising 7023 MRI images of the human brain, classified into four categories: glioma, meningioma, no tumor, and pituitary. This dataset is an amalgamation of data from three sources: figshare, the SARTAJ dataset, and Br35H. The images labeled as ‘no tumor’ were sourced from the Br35H dataset. Given concerns about the accuracy of glioma classification in the SARTAJ dataset, these images were replaced with those from figshare to ensure the integrity of the dataset. Table shows the dataset description while Fig. represents the dataset distribution.

Table 3. Dataset description

Fig. 2. Dataset distribution

Each image in the dataset underwent a series of preprocessing steps. This included augmenting the images to improve the model’s ability to generalize and learn from a more diverse range of data representations. The augmentation techniques included adjusting brightness and contrast levels randomly within specified ranges. The images were then resized to a uniform size of 128 × 128 pixels to ensure consistency in input data for the model which can be seen in Fig..

Fig. 3. Sample images from the dataset

Augmentation

In a bid to counteract overfitting and bolster the model’s capacity to generalize beyond the training set, augmentation techniques were employed. These techniques introduced random modifications to the images’ brightness and contrast, mimicking the variability often encountered in real-world medical imaging scenarios. By infusing diversity into the dataset through augmentation, the model is better equipped to adapt to varying image characteristics during training, potentially enhancing its ability to make accurate predictions on unseen data. In Fig. images after augmentation can be seen.

Fig. 4. Images after augmentation

The preprocessing steps undertaken in this study—augmentation, normalization, and resizing—significantly contribute to enhancing the dataset’s quality and preparing the images for efficient utilization within the CNN model. Augmentation broadens the dataset’s variability, normalization standardizes pixel values for effective model training, and resizing ensures uniform input dimensions, collectively aiding in building a robust and reliable model for brain tumor classification.

Moreover, the strategic curation and refinement of the glioma class within the dataset underscore the study’s commitment to data quality and diversity, crucial factors influencing the CNN model’s performance and its potential applicability in real-world scenarios. This comprehensive dataset, augmented and preprocessed to optimize its utility for model training, lays a solid foundation for the subsequent phases of the study, enabling the development of an accurate and adaptable brain tumor classification model.

B. Convolutional neural network (CNN) model architecture

The core of our methodology is the implementation of a Convolutional Neural Network (CNN) which can be seen in Eq., specifically leveraging the VGG16 model architecture. The VGG16 model, a product of the Visual Graphics Group (VGG) at the University of Oxford, has garnered acclaim for its prowess in image recognition tasks. Its architecture comprises a sequence of convolutional layers, interspersed with max-pooling layers, culminating in a series of fully connected layers. Trained on the ImageNet dataset, it gained popularity due to its ability to discern intricate features within images, making it an ideal choice for various computer vision applications.

Adaptation for brain tumor classification: In this research, the VGG16 model which can be observed in Eq. was adapted to cater specifically to the task of brain tumor classification. The original top layers, responsible for ImageNet’s classification into a thousand categories, were excised to tailor the architecture to the four distinct categories pertinent to brain tumors: glioma, meningioma, no tumor, and pituitary.

C. Model architecture

In preparing MRI images for compatibility with the VGG16 model, we implemented a series of preprocessing steps designed to optimize input data quality and consistency. This included resizing images to 224 × 224 pixels, the standard input size for VGG16, and applying a normalization process to scale pixel values to a range that matches the original VGG16 training data. To adapt the VGG16 architecture for the specialized task of brain tumor classification from MRI images, we introduced modifications that included fine-tuning the filter sizes in convolutional layers to better capture the nuances of MRI textures and adding additional dropout layers to prevent overfitting. Furthermore, our transfer learning strategy involved the utilization of pre-trained weights from the ImageNet dataset, leveraging the model’s existing feature extraction capabilities. This approach was complemented by fine-tuning the top layers of the model to align with our specific classification task, allowing the network to adjust to the distinct characteristics of brain tumors. The output layer of the model was reconfigured to support multi-class classification, replacing the original 1000-class output with a new layer designed to distinguish between four tumor categories: glioma, meningioma, no tumor, and pituitary. This layer employs a SoftMax activation function to output probabilities across these four categories, ensuring the model’s predictions align with the classification requirements of our study. Together, these tailored preprocessing steps, architectural modifications, and strategic application of transfer learning empower our CNN model to effectively classify brain tumors from MRI images with enhanced accuracy and generalizability.

The architecture of the CNN model, employing the modified VGG16 framework, embodies a structured hierarchy of layers meticulously designed for the intricate task of brain tumor classification. Leveraging the strengths of the VGG16 base model while customizing it to the specific requirements of our dataset, this architecture serves as a robust foundation for the subsequent phases of model training and evaluation, aiming to accurately classify brain tumors into distinct categories for enhanced clinical diagnosis and prognosis.

D. Federated learning implementation

To address data privacy concerns and improve model robustness, we adopted a federated learning approach. In this approach, the model training is decentralized, with multiple clients (in this case, 10) training models on their subsets of data. This method ensures that sensitive medical data does not leave its original location, preserving patient privacy.

The federated learning process involved the following steps:

Federated learning offers a robust and privacy-preserving paradigm for advancing brain tumor classification models. By distributing model training across multiple clients while maintaining data localization, this approach addresses data privacy concerns and augments model robustness through diverse datasets. The iterative refinement of the global model via model aggregation integrates insights from varied clinical contexts, paving the way for more accurate and adaptable brain tumor classification models with heightened privacy safeguards. The several steps that an image goes through during the classification can be observed in Fig..

Fig. 5. Different steps of image processing

E. Training and evaluation

The training process involved feeding the CNN model with batches of images, with a specified batch size and number of epochs. The model’s performance was evaluated using standard metrics, including accuracy, precision, recall, and F1-score.

In addressing the intricacies of federated learning within our brain tumor classification model, we acknowledge the inherent challenge of increased communication overhead that this distributed training approach entails. Federated learning necessitates frequent exchanges of model updates between the client and the central server, which can significantly strain network resources. To mitigate this overhead, we have employed strategies such as model compression techniques and sparsification, which reduce the size of the model updates being transmitted without compromising the integrity of the training process. A separate set of images, not used during the training phase, constituted the test dataset. This dataset was critical for assessing the model’s ability to generalize and accurately classify unseen data.

Training batches

Fig. 6. Training loss

Results and discussions

When assessing the effectiveness of a brain tumor classification model, a range of evaluation metrics is employed to gain comprehensive insights into its performance. These metrics serve as pivotal benchmarks to gauge the model’s accuracy, its ability to correctly identify different tumor types, and its overall efficacy in handling the classification task.

In brain tumor classification, these metrics play a pivotal role in understanding the model’s ability to discern between different tumor types. For instance, precision would reveal how accurately the model identifies a specific tumor type among all the instances it classified as that type. Meanwhile, recall would emphasize how well the model detects all instances of a particular tumor type among the total instances of that type in the dataset.

In the next section, we will present and discuss the results obtained from this methodology.

The results obtained from the federated learning-based CNN model demonstrate its exceptional capability in accurately classifying brain tumors from MRI images. The model’s performance was assessed through various metrics including precision, recall, F1-score, and overall accuracy, supported by a detailed analysis using a confusion matrix.

A. Model performance

The performance of the federated learning-based CNN model was rigorously evaluated using a comprehensive classification report, accuracy scores, and a confusion matrix. The following are the key findings:

Table 4. Classification report

Fig. 7. Classification report

Fig. 8. Confusion matrix

B. Analysis of results

The results from the model’s performance evaluation reveal several key insights:

In conclusion, the proposed federated learning-based CNN model has proven highly effective in the classification of brain tumors using MRI images. Its high accuracy, precision, and recall make it a promising tool for aiding medical professionals in the diagnosis and treatment planning of brain tumors. The following section will discuss these results in the context of existing methodologies and explore their implications in the field of medical imaging.

C. Comparison with existing methods

The federated learning-based CNN model marks a significant advancement over traditional and existing deep learning methods in brain tumor classification. Traditional approaches, reliant on manual interpretation of MRI images, are time-consuming and subject to human error. Even with the integration of conventional machine learning techniques, these methods often lack the robustness and adaptability required for precise tumor classification.

In contrast, existing deep learning methods, while more accurate than manual interpretation, frequently encounter challenges related to data privacy, model generalization, and dependency on large, well-annotated datasets. The proposed federated learning model addresses these concerns effectively. Its ability to achieve a high accuracy rate of 98%, with substantial precision and recall across all tumor types, sets it apart from previous deep learning approaches. Moreover, the decentralized nature of federated learning ensures data privacy and diversity, contributing to the model’s robust performance across various datasets. In the below table the comparison with the existing model is given in Table:

Table 5. Comparison with existing methodologies

D. Challenges and limitations

Despite its successes, the study faces several challenges and limitations:

The generalization of findings to diverse patient populations remains a limitation. The study’s results, while promising, require further validation in cohorts that encompass a wider range of demographic, geographic, and clinical characteristics. Such validation studies are essential to confirm the model’s applicability and effectiveness across different settings and populations, ensuring that the benefits of federated learning for brain tumor classification can be realized on a global scale.

E. Future directions

To further enhance the model’s efficacy and applicability, several future directions are suggested:

In summary, while the federated learning-based CNN model presents a significant improvement in brain tumor classification from MRI images, ongoing research and development are essential to address its current limitations and to explore its full potential in the broader context of medical imaging.

Conclusion

This study has successfully developed and evaluated a federated learning-based Convolutional Neural Network model for the classification of brain tumors using MRI images. The main findings include the model’s remarkable accuracy rate of 98%, along with high precision and recall across all tumor types: glioma, meningioma, no tumor, and pituitary. These results signify a notable improvement over traditional and existing deep learning methodologies, primarily due to the incorporation of federated learning, which enhances data privacy and model generalization. The study has effectively demonstrated the feasibility and efficacy of using advanced machine learning techniques in the critical domain of medical imaging.

The implications of this research are far-reaching and transformative for the field of medical diagnostics and patient care. The high accuracy and efficiency of the model in classifying brain tumors can significantly aid radiologists and oncologists in making more informed and quicker diagnostic decisions, potentially leading to earlier and more effective treatment plans. This advancement is especially crucial in brain tumor cases, where early detection and accurate classification can markedly influence patient outcomes.

Moreover, the successful application of federated learning in this context opens new avenues for medical data analysis while respecting patient privacy and data security concerns. The model’s approach can be extended to other types of medical imaging tasks, paving the way for broader applications in healthcare diagnostics. It also sets a precedent for future research to explore and refine machine learning models, making them more accessible and practical for clinical use.

In conclusion, this research contributes significantly to the intersection of artificial intelligence and healthcare, showcasing the potential of machine learning to revolutionize medical diagnostics and enhance patient care. However, we recognize the importance of further validation on larger and more diverse datasets to assess the generalizability and robustness of our approach in real-world clinical settings. Additionally, we advocate for exploring extensions of the federated learning framework in other medical imaging tasks beyond brain tumor classification, paving the way for advancements in privacy-preserving AI-driven healthcare solutions. We believe that continued research in this direction will contribute significantly to improving the efficiency, accessibility, and effectiveness of medical imaging technologies, ultimately benefiting patient care and outcomes.

Consent for publication

Not applicable as the work is carried out on publicly available dataset.