Deriving Efficient 3D U-Net Based Segmented Anomaly Detection and Classification in 3D MRI Images Using ConvLSTM Model and Shuffled Frog Leaping Algorithm

Abstract

Anomaly detection relates to the difficulty of detecting anomalous inputs from seen samples of standard data. Despite recent advances in deep learning for detecting visual anomalies, these algorithms are still incapable of deciphering complex images like those encountered in the medical business. Medical imaging is essential for diagnosing and treating a vast array of conditions, including anomalies detected in 3D Magnetic Resonance Imaging (MRI) data. The dependable and accurate detection and classification of anomalies within volumetric MRI scans remain challenging for the medical industry. This study presents a novel method for combining three potent techniques: 3D U-Net for segmentation, the Convolutional Long Short-Term Memory (ConvLSTM) model for temporal processing, and the Shuffled Frog Leaping Algorithm (SFLA) for optimization. Utilizing the 3D U-Net architecture, the proposed method efficiently segments regions of interest in 3D MRI volumes, allowing for precise anomaly localization. The ConvLSTM model incorporates temporal dependencies between successive MRI slices, enhancing the accuracy of detecting and classifying dynamic anomalies. Combining the SFLA as a metaheuristic optimization instrument significantly improves the efficiency and effectiveness of the proposed framework. The SFLA optimizes the network's hyperparameters, improving convergence and lowering the danger of training being trapped in local optima. In order to calculate the Anomaly Class Weight, the output layer neurons are built to estimate various Feature Distribution Similarity values for various characteristics. Extensive experiments were conducted on a large dataset of 3D MRI scans with various defects to assess the efficacy of the proposed technique. According to the results, the proposed method outperforms standard anomaly detection and categorization procedures. The method accomplishes cutting-edge precision, sensitivity, and specificity, surpassing existing approaches by a wide margin.