Some Metaheuristic Algorithms to Design Deep Neural Networks for Medical Image Detection

Abstract

Medical image analysis plays a critical role in modern healthcare for treatment plan ning, diagnosis, and disease prediction, including brain tumors. It encompasses various modalities, including X-ray, Computed Tomography (CT), and Magnetic Resonance Imag ing (MRI), each with unique strengths and applications. Analyzing these complex im ages pose challenges due to noise, and image quality variability, making them prone to errors and heavily reliant on the knowledge and experience of physicians. Medical imag ing tasks, such as denoising, aim to improve image quality for precise diagnosis, while accurate segmentation enables quantitative analysis and visualization of anatomical abnor malities, with early detection of brain tumors crucial for timely intervention and improved patient outcomes. Deep Neural Networks (DNNs) have shown promising results in medical image analysis. However, manually designing DNN models is challenging, tedious, time consuming, and requires domain-specific knowledge. The increasing number of available training techniques adds complexity to finding the optimal structure, and selecting the suit able hyperparameters for a given task often entails multiple trial-and-error iterations. To address these challenges, Neural Architecture Search (NAS) has emerged as a promising solution, which automates the design of DNNs for specific tasks. Nevertheless, NAS meth ods need further optimization in designing a search space, constructing a DNN from search space (encoding), and exploring different search strategies for specific tasks. The Meta heuristic Algorithm (MA) based methods in NAS have gained traction for automating the DNN architecture design process. This thesis proposes automatically designing flexible and efficient DNN architectures and hyperparameters using MAs. Integrating metaheuris tic optimization with deep learning can lead to adaptable and practical solutions for medical image analysis. Flexible search spaces, advanced techniques, and consideration of compu tational resources contribute to developing practical solutions. The proposed metaheuristic optimization framework has the potential to revolutionize medical image analysis, enhanc ing patient care by enabling better diagnosis, treatment planning, and research in medical imaging. The main objectives of this thesis include: (i) To design a metaheuristic block-based iv deep neural network for medical image denoising, (ii) To develop a metaheuristic-based modified U-shaped network for 2D medical image segmentation with denoising, (iii) To develop a metaheuristic based encoder-decoder model for 3D medical image segmentation, and (iv) To design a multi-objective metaheuristic model for detecting brain tumors in 3D medical images. In this thesis, to achieve the abovementioned objectives, we proposed some metaheuristic-based approaches for medical image analysis tasks, including denoising, seg mentation, and brain tumor detection. Firstly, a metaheuristic block-based deep neural net work is designed for medical image denoising. The denoising performance is enhanced by utilizing the Differential Evolution (DE) algorithm, which facilitates the exploration of various combinations of network components and hyperparameters within the specified search space. Secondly, a metaheuristic-based modified U-shaped network is developed for 2D medical image segmentation with denoising. A modified U-shaped architecture is used with a flexible search space that allows the optimization of individual blocks. Further more, attention blocks are incorporated to enhance segmentation accuracy. The Teaching Learning-Based Optimization (BTLBO) algorithm is used for optimization, resulting in im proved segmentation performance. Thirdly, a metaheuristic-based encoder-decoder model is developed for 3D medical image segmentation. A powerful search space is constructed to optimize the network blocks and training parameters. The Chameleon Search Algorithm (CSA) explores the search space to improve the segmentation performance. Lastly, the third objective is extended to brain tumor detection using a multi-objective optimization ap proach to optimize detection performance and network size. The search space is expanded to include various blocks and parameters. The Multi-objective Iterative Teaching-Learning Based Optimization (MO-ITLBO) algorithm is utilized to identify optimal block structures and training parameters. The experimental results of this research demonstrate the effec tiveness of metaheuristic optimization techniques in enhancing various tasks of medical image analysis. The proposed models outperform existing methods, offering improved de noising, segmentation, and brain tumor detection performance. These advancements can revolutionize medical image analysis, leading to better patient care, diagnosis, and treat ment planning in medical imaging.