<p>Artificial intelligence (AI) has become a central component of medical image analysis, leading to the emergence of diverse architectural paradigms with distinct methodological characteristics and practical implications. However, their relative strengths, limitations, and applicability across clinical tasks remain insufficiently synthesized. This systematic review follows the PRISMA guidelines and analyzes 123 peer reviewed studies published between 2023 and 2026 in Scopus and Web of Science. It provides a structured comparison of five major AI paradigms: Vision Transformers (ViT), Vision Language Models (VLM), Large Language Models (LLM), Mamba based Selective State Space Models (SSM), and fuzzy logic based approaches, across a range of imaging modalities and clinical applications. The findings indicate that ViT based models generally demonstrate improved performance compared to convolutional architectures, particularly in segmentation tasks, although their effectiveness depends on data availability and computational resources. Self supervised learning strategies contribute to reducing reliance on annotated data, while Mamba based SSMs offer computational advantages for modeling long range dependencies. In addition, hybrid fuzzy deep learning approaches provide a mechanism to incorporate uncertainty into decision making, leading to performance improvements in specific scenarios. Vision language models show promising results in multimodal interpretation tasks; however, their clinical adoption remains limited due to challenges related to robustness, interpretability, and regulatory validation. The results suggest that no single paradigm consistently outperforms others across all tasks. Instead, the selection of appropriate models should be guided by the clinical objective, data characteristics, and deployment constraints. These findings highlight the importance of integrative and task specific approaches for advancing AI driven medical imaging.</p>