Content-Based Brain Magnetic Resonance Image Retrieval and Classification With the Proposed Deep Learning and Tissue-Based System

Abstract

The exponential growth in the size of databases due to technological advancements has led to challenges in locating and accessing specific components of the data. While deep learning and other machine learning architectures have shown promise in retrieving data components, their efficacy is more pronounced when addressing disease cohorts. Contrarily, this effectiveness diminishes when accessing large datasets. This study focuses on the analysis of brain magnetic resonance imaging (MRI) images and, specifically, to differentiate between benign and malignant lesions associated with Alzheimer's disease, multiple sclerosis (MS), and intracranial regions, all of which are medically significant with distinct treatment modalities. A hybrid model was first devised to facilitate image retrieval by employing a pre-trained EfficientNet-b0 and local binary pattern (LBP) for feature extraction. These extracted features were then amalgamated to encompass diverse aspects of each image. To improve model performance, redundant features were pruned using the minimum redundancy maximum relevance (mRMR) technique. As a result, the proposed model demonstrated efficacy in analyzing a diverse dataset encompassing three distinct diseases and eight unique classes. Notably, existing machine architectures already published in the literature have struggled to achieve comparable success rates in discerning such closely related yet distinct disease groups. Our study underscores the challenge posed by increasing class diversity on the performance of deep learning architectures and obtained an accuracy of 98.9% in classifying three diseases and eight unique classes. As a result, the same model was used as the base in both the classification and CBIR processes for MRI detection, yielding competitive results when compared with the literature and other models.