The Multiple Brain Tumor with Modified DenseNet121 Architecture Using Brain MRI Images

Abstract

Brain tumors are capable of developing in individuals of all ages and can originate from brain tissue in various shapes and sizes. As a result, it is critical to quickly identify patients in order to expedite treatment. Magnetic Resonance Imaging (MRI) of the brain is an appropriate technique for identifying chronic conditions, including tumors. Deep learning methodologies have suggested numerous medical analysis strategies for health monitoring and brain tumor identification. This study used a modified version of DenseNet121 to accurately categorize three different forms of brain tumors: meningioma, pituitary, and glioma. Following the last transition layer, the DenseNet121 modification adds DropOut and GlobalAveragePooling layers. We determine the optimal hyperparameters that yield the highest performance by comparing several factors, including dropout, epoch, optimizer, and activation function. Evaluation of classification performance involves a comparison between Basic CNN and Basic DenseNet. Results of the analysis show that the modified DenseNet121 model works best with the following ideal hyperparameters: ADAM optimizer, Softmax activation, 150 epochs of training, and an 0.8 dropout rate. The performance results show an accuracy value of 0.9782, exceeding previous research findings.

Keywords

• Convolutional neural network
• Classification Algorithms
• Hyperparameter Optimization
• Digital image Processing
• Deep Learning, CNN, Multi Layer Perceptron, Konvolusi, Tensorflow.