Expert-guided extraction of relevant informations : Application to medical pathology detection

Abstract

Magnetic Resonance Imaging (MRI) brain tumor identification and classification are costly and time-consuming due to tumor complexity and reliance on radiologist expertise. To overcome these challenges, automating the process is essential. This thesis leverages the power of deep learning for brain tumor analysis, presenting two key contributions. In the first contribution, we introduce an efficient model titled "Deep Rule-Based Classifier using Bank of Binarized Statistical Image Features (DRB-BBSIF)". This approach addresses the limitations of conventional MRI brain tumor diagnosis by offering a model that improves classification performance while reducing the complexity of the diagnostic process. The model explores the BSIF image descriptor for the feature extraction phase, Furthermore, to enhance its performance, we have constructed a Bank-BSIF, which is founded by the best parameters of BSIF filters. For the classification phase, we employed a deep rule-based (DRB) classifier. The DRB classifier functions through a self-organized set of IF-THEN fuzzy rules, guided by . These fuzzy rules, generated by the DRB classifier, serve as the classifier's core decision-making mechanism. The second contribution titled “MRI Brain Tumor Identification and Classification using Deep Learning Techniques” focuses on the synergistic integration of deep learning and rule- based classification. We propose a novel, simple, and automatic DRB-based scheme for MRI brain tumor classification. This model leverages the power of deep learning for feature extraction and combines it with the effectiveness of DRB for classification. The framework consists of three stages: preprocessing, feature extraction, and classification. Feature extraction utilizes deep learning networks like AlexNet, VGG-16, ResNet-50, and ResNet- 18 to extract features from the MRI images. A DRB classifier then utilizes these deep features for classification.