Implementación de un modelo generativo de Deep Learning para re-muestrear imágenes de resonancia magnética

Abstract

DOCUMENT IN THE PROCESS OF INTELLECTUAL PROPERTY. This study develops a generative model based on Generative Adversarial Networks (GANs) to generate intermediate slices in three-dimensional Magnetic Resonance Imaging (3D MRI), hypothesizing that GANs surpass traditional interpolation limitations by preserving anatomical structures. The Pix2Pix architecture was implemented, trained on a dataset of 93 subjects (38 healthy, 55 Parkinson’s patients), preprocessed through bias field correction and intensity normalization. Slices were segmented into odd and even pairs, combining a U-Net generator with residual blocks and a convolutional discriminator. The final model (V5) achieved a Structural Similarity Index (SSIM) of 0.8469 and a Peak Signal-to-Noise Ratio (PSNR) of 30.05 dB, demonstrating qualitative superiority over conventional methods. The algorithm was integrated into a web application with user/administrator roles for uploading, processing, and visualizing images. It is concluded that GANs enhance spatial continuity in 3D MRI by generating coherent intermediate slices, reducing artifacts, and preserving critical anatomical details. Clinical validation and dataset expansion with diverse demographic and pathological cases are recommended to ensure robustness in real-world medical applications. Keywords: generative adversarial networks, three-dimensional interpolation, medical imaging, synthetic image generation.