AI-driven medical dataset quality enhancement improves statistical fidelity and downstream utility in select modalitiesAI tools show promise in improving medical data quality

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.