Thesis:

Title : Proposing a Method Based on Artificial Neural Networks for Improved Edge Detection in MRI Images

Abstract: Medical image processing, particularly the precise detection of anatomical boundaries, is considered a fundamental step in the development of computer-aided diagnosis systems and clinical analyses. Magnetic resonance imaging -MRI- faces complex challenges in the edge detection process due to subtle differences in the intensity of soft tissues and the presence of inherent noise in imaging systems. Classical and conventional methods, such as derivative-based operators and multi-stage thresholding algorithms, typically encounter boundary discontinuities when confronting these structural complexities, or produce numerous false and distracting edges within internal tissue regions due to over-segmentation. To overcome these fundamental limitations, the current research presents a novel and intelligent approach based on a Multi-Layer Perceptron -MLP- neural network for the precise extraction of edges in brain MRI images. In this structure, the model is fed by utilizing spatial features and the targeted extraction of directional intensity derivatives to learn authentic edge patterns in the absence of heavy computational burdens. For a comprehensive evaluation and to prove the system's efficiency, the outputs of the proposed architecture were analyzed in comparison with classical methods using a set of image quality metrics including MSE, PSNR, and SSIM, as well as machine learning criteria such as Precision, Recall, and F1-Score. The results of visual and analytical evaluations proved that the developed architecture, instead of memorizing the noise present in the training labels, has exhibited a behavior based on intelligent filtering. The proposed network successfully extracted the main anatomical boundaries with complete continuity and standard thickness, and simultaneously cleared the internal tissue spaces of any edges lacking diagnostic value. In addition to superior visual quality and stability in the convergence process, examining the algorithm's execution time showed that this model, recording a processing time of approximately a few hundredths of a second for each slice, has the capability to be deployed in real-time processing systems. In a general conclusion, this thesis demonstrates that utilizing optimized artificial neural network architectures can be a stable, fast, and highly reliable solution for enhancing image processing systems in clinical environments.

Keywords: Edge detection, Magnetic Resonance Imaging (MRI), Multilayer Perceptron (MLP) neural network, medica