Zeitschrift für Hirnforschung

Detection of Brain Tumor Using Deep Convolutional Network

Abstract

Srishti Singh, Dhriti Sood, Soumya Tyagi,Shubhi Jadaun, Nathi Ram Chauhan5

Introduction: More than 28,000 people are diagnosed with brain tumor every year in India, out of which 80% are cancerous and more than 24,000 people die due to delayed treatment of tumor and cancer identification. This work is focused on "Brain Tumor Detection Using Deep Convolutional Neural Network” and it proposes a method for accurately detecting brain tumours with deep Convolutional Neural Networks (CNNs), eliminating the need of a radiologist to confirm the presence and identifying the type of tumor, thereby putting more focus on immediate treatment and cancer identification using biopsy and reducing the duration of the tumor detection process. This work highlights the importance of early and accurate detection of tumors in brain, as it plays critical role in the successful treatment and management of a brain cancer. It is mentioned that traditional methods for brain tumor detection, such as manual segmentation and feature extraction are timeconsuming and subject to human error. Hence, deep learning-based approach using 23 layers CNN architecture and transfer learning based VGG16 architecture has been proposed, which have shown remarkable success in a various image recognition tasks.

Methodology: The methodology used in the research, involves several steps. First step involves collection of brain MRI (Magnetic Resonance Imaging) data from a publicly available dataset- Harvard Medical Dataset and Figshare. The data by resizing the images to a consistent resolution and normalizing the pixel values. Then, the dataset is split into training, and testing sets, in the ratio of 7:3 fora training and evaluating the proposed CNN models. Ab CNN architecture is designed, which consists of multiple convolutional and pooling layers followed by a fully connected layer. Rectified Linear Unit (ReLU) activation functions are used to introduce non-linearity and batch normalization to improve the training stability. Dropout regularization is employed to prevent overfitting, which is a common issue in deep learning models.

Discussion: CNN models are trained using the training set and optimized using Stochastic Gradient Descent (SGD) with a categorical cross-entropy loss function. Experimentations with different hyper parameters, such as learning rate and batch size, were carried out to find the optimal settings for the models. Data augmentation techniques were performed, such as rotation, flipping, and scaling, to increase the diversity of training data and improve the model’s generalization ability. Once the training was completed, the CNN model was evaluated on the testing sets. Various performance metrics were reported, such as accuracy, precision, recall, and F1-score, true positive rate and true negative rate to assess the effectiveness of the models in detecting brain tumors accurately. The results were compared with existing methods in the literature and it was observed that the proposed CNN models outperform all of them in terms of accuracy and other performance measures.

Conclusion: Furthermore, additional experiments are carried out to analyze the robustness of the CNN models against different types of brain tumors, such as benign and malignant tumors, as well as different tumor sizes. Ablation studies are performed to investigate the impact of different components of the CNN architecture on the model's performance. In addition, Different kernel sizes, which refer to the width height of the filter mask here, are integrated with the model to extract the complex features from the MRI images to make the model more robust and adaptive. The radiologist uses different experimental procedures for diagnosing brain tumors, including biopsy, Cerebrospinal Fluid (CSF) analysis, and X-ray analysis. The biopsy process introduces many risks including inflammation and severe bleeding. It also has just 49.1% accuracy.

CSF Analysis, similar to biopsy, it introduces many risks including bleeding from the incision site to the bloodstream and perhaps an allergic reaction after the treatment. Similarly, using X-rays on the skull can lead to an increase in the risk of cancer due to the radiation. The results of the research show that the proposed CNN-based approach achieves high accuracy in detecting brain tumors, with promising performance metrics across different types and sizes of tumors. The implications of the findings, includes the potential clinical applications of the approach in real-world scenarios fora assisting radiologists in accurate brain tumor diagnosis. The limitation of the study, is that only a single dataset is used for testing, however testing should have been performed on a diverse dataset of real clinical images.