Establishment of High-Precision Ultrasound Diagnosis Methods Based on the Introduction of Deep Learning

Ultrasound imaging is widely used owing to its affordability, radiation-free, and non-invasive advantages. However, limitations stemming from operator dependence and artifacts have been noted. To address these issues, deep learning (DL) is increasingly being introduced. In oncology and cardiology, DL-equipped devices are transitioning to clinical use following approval. Nevertheless, DL faces challenges such as generalization, safety, and operational burden, making strategic implementation essential to maximize patient benefit. Existing reviews often list individual technologies but lack evaluation frameworks tailored to clinical implementation. Therefore, this review (i) organizes and formalizes limitations specific to ultrasound diagnosis, (ii) explains the latest DL methods addressing these limitations in terms of principles, implementation, and evaluation metrics, and (iii) examines recent clinical applications, including approved devices, supported by evidence, demonstrating that DL possesses substantial utility beyond the research stage for improving clinical workflows. It also critically evaluates remaining challenges, presents evaluation criteria to aid implementation, and identifies future research challenges.