AI-driven quality enhancement of medical datasets: a closed-loop framework of synthesis, annotation, quality control,...

The performance ceiling of medical AI is largely determined by the quality of its training datasets, yet traditional manual collection and annotation face a triple bottleneck of high cost, embedded bias, and limited scalability. This review takes "quality enhancement" as its unifying axis, integrating the four pillars of AI-driven medical dataset quality enhancement-synthetic data generation, intelligent annotation, automated quality control, and integration pipelines-and proposes a four-pillar integration framework as its core contribution. The framework organizes the four pillars into a closed loop: synthesis and annotation are coupled through dual-track feedback (a quantity track that replenishes data and a strategy track that adjusts generation strategy), while annotation and quality control are coupled through bidirectional verification. Beyond this forward enhancement loop, the framework further employs a bias amplification chain (as a heuristic insight) and the privacy-utility tradeoff as two tension lines, and incorporates Metrics Reloaded, FAIR, ALCOA+, and YY/T 1833 as sub-layer evaluation dimensions in a complementary manner. The review shows that AI-based enhancement methods have approached real data in statistical fidelity and downstream utility for select modalities, tasks, and evaluation settings, but systematic gaps remain in privacy, fairness, and clinical validity; single-pillar optimization is hypothesized, on the basis of indirect segmental evidence, to be insufficient to improve end-to-end quality, and closed-loop synergy is proposed as a testable proposition rather than an established requirement; integration pipelines are already usable for single-disease, single-modality settings, whereas cross-institutional, cross-modal, and cross-disease generalization remains the principal barrier, with federated construction and foundation-model-driven approaches as promising directions.