<p>Segmentation of medical images serves as the indispensable cornerstone of modern diagnostic and therapeutic frameworks, providing the precision required to delineate complex anatomical structures and pathological entities across modalities such as MRI, CT, and ultrasound. The advent of Deep Learning (DL) has fundamentally shifted this landscape, with Convolutional Neural Networks (CNNs) and Transformer architectures achieving unprecedented accuracy. However, the transition from algorithmic innovation to clinical translation is hindered by the “Medical Data Triad” of scarcity, heterogeneity, and opacity. This review presents a comprehensive taxonomic synthesis of the state-of-the-art in “Hybrid Intelligence”—defined here as the synergistic integration of localized inductive biases from CNNs with the global relational modelling of Transformers—to overcome these persistent barriers. Unlike prior surveys, this work provides a critical meta-analysis of “Data-Efficient” methods, including Self-Supervised Learning (SSL), Federated DL, and synthetic data generation via Diffusion models, specifically evaluating their efficacy in mitigating data scarcity while preserving patient privacy. We perform a rigorous cross-comparative assessment of benchmark performance (e.g. BraTS, Synapse, ACDC) against operational constraints such as computational complexity and real-time clinical feasibility. Finally, we bridge the “Technical-Clinical Divide” by mapping regulatory pathways (FDA/EMA) and identifying implementation science gaps, offering a strategic roadmap for the next generation of AI-enhanced, interpretable medical imaging systems.</p>