Hapsari Peni Agustin Tjahyaningtijas, Lusia Rakhmawati, Pradini Puspitaningayu, Syaqila Sal Sabila, Laras Suciningtyas, Luthvi Ramasandhika Yustiadi, Ari Kusumaningsih
Brain tumors can occur in people of all age groups and can originate from various types and sizes of brain tissue. Prompt patient detection is critical for accelerating therapy. Magnetic Resonance Imaging (MRI) is a highly effective technique for detecting chronic disorders like brain tumors. While deep learning techniques dominate medical analytics for health monitoring and brain tumor identification, older machine learning approaches often receive less attention despite their comparable performance. The classification accuracy achieved by these methods is inadequate because they rely solely on texture-based features and employ techniques that fail to capture some key aspects of the images. Additionally, these methods were prone to overfitting. This study substitutes traditional handcrafted feature extraction methods with automated feature extraction using a Convolutional Neural Network (CNN).In this research, we used DenseNet because of its ability to enhance information flow across layers, minimize parameter numbers, and substantially mitigate vanishing gradients. For the classification, we employed XGBoost to enhance classification performance based on the high-level features extracted by DenseNet. We include a second-to-last layer (Global average Pooling2D) in the suggested architecture to extract features before the dense classification layer and then use the extracted features for XGBoost training. We use the DenseNet121 model as the foundational architecture for learning features. We determine the optimal hyperparameters for the CNN architecture, such as optimizers, activation functions, and epochs, in order to maximize performance. The classification performance evaluation compares the basic CNN and hybrid CNN+SVM models. The analysis indicates that the hybrid CNN+XGB model exhibits exceptional performance, achieving an accuracy of 0.9827 on brain tumor image classification using T1 MRI modality Chandrabaga Clinic and Nursing Home Dataset, exceeding the results of previous investigations. This success is attributed to utilizing optimal hyperparameters, including the ADAM optimizer, softmax activation, and 150 training epochs. © (2025), (Intelligent Network and Systems Society). All rights reserved.
Department of Electrical Engineering, Universitas Negeri Surabaya, Indonesia; Department of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Indonesia; Department of Informatic Engineering, Universitas Trunojoyo Madura, Indonesia
