Comparación entre DCGAN y PGGAN para la generación de imágenes de resonancia magnética en demencia frontotemporal

Abstract

Introduction: The advancement of artificial intelligence (AI) in medicine has exposed critical challenges, particularly the shortage of high-quality medical data and associated ethical concerns. Frontotemporal dementia (FTD) is a group of neurodegenerative conditions where early diagnosis is crucial. However, the development of AI-assisted diagnostic systems faces challenges due to the scarcity of labeled magnetic resonance imaging (MRI) data. Objective: To develop, compare, and evaluate two Generative Adversarial Network (GAN) models, Deep Convolutional GAN (DCGAN) and Progressive Growing GAN (PGGAN), for generating synthetic brain MRI slices, assessing image-level similarity and visual plausibility. Methods: We trained both models on a limited dataset of FTD brain MRIs. Our evaluation encompassed quantitative and qualitative assessments. We used three metrics for quantitative analysis: Structural Similarity Index (SSIM), Normalized Mutual Information (NMI), and Peak Signal-to-Noise Ratio (PSNR). Results: Both models generated synthetic MRIs. SSIM values were similar (≈0.215) with no practical difference; PGGAN surpassed DCGAN in NMI and PSNR (0.976 versus 0.798; 14.900 versus 12.426). Qualitative analysis corroborated these findings, with PGGAN producing images featuring enhanced detail. Conclusion: This research contributes to the field of synthetic medical image generation, offering a potential solution to data scarcity while maintaining patient privacy. The observed performance of PGGAN holds promise for augmenting limited datasets in neurological research. Furthermore, this study lays the groundwork for developing robust AI-assisted diagnostic tools in neurology, particularly for conditions like FTD, where extensive, diverse datasets are crucial yet challenging. Keywords: Artificial intelligence; frontotemporal dementia; magnetic resonance imaging; deep learning (Source: MeSH NLM)

Ähnliche Arbeiten