Fine Tuned Deep Learning Model for Multi-Class Brain Tumor Detection and Classification

Abstract

Depression is a serious mental disease which has affected millions of people all over the world, but it is very difficult to detect in early-stage. Traditional diagnosis is Mesh Keywords based on Image processing plays a significant role in neurologists' medical diagnoses in the medical field. MRI(magnetic resonance imaging) is certainly the keystone of brain tumor imaging among all kinds of imaging, playing a crucial role in all phases of patient management, starting from medical diagnosis, through therapy preparation, to treatment reaction and reoccurrence assessment. This extensive study explores the capacity of innovative deep-learning strategies for the critical task of categorizing brain tumors from magnetic resonance imaging (MRI) scans. The methodology incorporates transfer learning, including the acquisition of a curated preprocessing with image grayscaling, data augmentation, and normalization to help with optimal model training. The research study utilizes a series of traditional machine learning approaches such as Decision Tree Classifier, Random Forest, K-Nearest Neighbors, Support Vector Machine(SVM), Logistic Regression, and Multi-Layer Perceptron (MLP), with accuracy respectively (75.5%, 81%, 76%,61%, 74%, and 61%) and sophisticated neural network architectures, with a particular concentration on the basic - CNN, Inception V3, customized - DenaseNet121, VGG16, ResNet50 designs, alongside an ingenious customized VGG16- Resnet50 hybrid model achieved accuracy respectively-(86.5%, 87%, 91%, 96.5%, 96%, and 98.5%). In order to recognize the most accurate and reliable method for clinicalapplication each model's performance was seriously assessed based on some performance metrics, including precision, recall, F1 score, Cohen's kappa statistic, sensitivity, andspecificity. In contrast, the deep learning designs displayed remarkable effectiveness on a diverse dataset of 40,000 MRI images covering four classes: glioma, meningioma, pituitary growth, and no tumor. The proposed hybrid approach demonstrated superior performance with a precision of 98.5%, outperforming the conventional machine- learning classifiers and standalone deep-learning models. This work highlights the capacity of hybrid architectures in supporting radiologists with trusted, automatic brain tumor detection systems.