Design of a Predictive Modeling System for MRI Brain Image Classification

An accurate identification and characterization of any potential abnormalities in the MRI brain image is required for efficient classification of brain cancer.Recent advancements by computerized diagnosis eradicate the possibility of radiologists making an incorrect diagnosis based on their skills to perceive and interpret data.This work proposed an efficient deep learning based Predictive Modeling System for Brain Cancer Classification (PMS-BCC).It consists of three modules: feature extraction module, feature selection module and classification module.To localize the patterns for brain cancer, PMS-BCC uses stacked convolution and pooling layers along with random skip connections.The main advantage of using this combination is that it allows the model to learn hierarchical and spatially invariant features efficiently, while also addressing common issues like vanishing gradients and overfitting.In the feature extraction module, the convolution layer is responsible for the extraction of locally relevant features, while the pooling layers minimize the feature dimension.In the subsequent module, an AntLion Optimization (ALO) is used to choose the optimal subset of features, and a neural network using Greedy Layerwise Training (GLT) is used in the classification module to do the classification.The gradient based optimization techniques suffer from local minima.ALO does not depend on gradients and can explore search space more effectively to avoid local minima compared to gradient based techniques and its variations.ALO has a better exploitation-exploration balance due to its specific hunting strategy than other metaheuristic algorithms.Results showed that the proposed PMS-BCC architecture with GLT improves the classification accuracy from 93.1% (convention training) to 98.6% for binary classification (normal/abnormal) and 93% (conventional training) to 98.2% for multiclass (normal/low-grade/high-grade) classification of MRI brain images obtained from REpository of Molecular BRAin Neoplasia DaTa (REMBRANDT) database.Though the proposed PMS-BCC system provides promising results, only the axial views of brain in the REMBRANDT database is used for the analysis.