人工智能在放射影像中的应用

近年来，人工智能技术的兴起为放射影像处理的相关应用提供了更加准确、更加高效的解决方案。人工智能技术的高速发展也在持续地为放射影像处理提供更实用、更具变革性的技术手段。人工智能技术多层次的特征提取模式显著提升了影像分析的精度，其端对端的学习模式极大提升了影像处理的速度，其多模态的学习能力则能提供更为全面的影像诊断结果。然而，人工智能技术的内在局限性也引起广泛的关注。模型对高质量标注数据的依赖、模型的泛化能力与可解释性仍是人工智能在应用过程中的主要挑战. 为了更好地厘清人工智能在放射影像处理领域的发展脉络，明确未来的研究重点，本文总结了近年来人工智能在放射影像诊断、配准、分割、生成与重建五个方面的代表性应用，并基于共引用频率分析方法选取、介绍了相关的奠基性工作，着重探究了人工智能在放射影像应用领域中普遍面临的挑战与最新的发展趋势。

Artificial intelligence (AI) techniques are thriving in the area of radiography, providing more accurate and more efficient solutions for processing and analyzing radiographic imaging, including X-ray, CT, MRI, ultrasound and nuclear imaging. As the rapid development of AI explosively provides more practical and transformative techniques, to empower more performant model designs and further hasten clinical deployments, it’s crucial to clarify the developing trajectory of the applications of AI in radiological image processing and analysis. The development of AI in radiography can be summarized in 6 stages, traditional rule-based optimization methods based on manual designed feature extraction algorithms, statistical machine learning, big data empowered deep learning, transformative innovations in deep learning networks, multimodal learning and multitask learning, and generalizability and interpretability learning in the era of foundations models. Along the six stages, this paper investigated the developing trajectories of the applications of AI in radiological image diagnosis, registration, segmentation, synthesis and reconstruction. Based on co-citation frequency analysis, we selectively introduced recent representative work and previous foundational work from these five application aspects. The main contribution of this paper is that we located prevailing challenges and recent advances. The successive births of transformative deep learning paradigms keep expanding more possible clinical scenarios, and the maturation of multimodal and multitask learning strategies and the rise of foundation models have provided a fundamental basis for precision medicine and personalized diagnosis, marking the start of a new era of comprehensive techniques integration and intelligent system design for radiological imaging processing and analyzing. In the meantime, although deep learning methods have significantly improved the accuracy, speed and comprehensiveness of radiological imaging processing and analysis, the reliance on high-quality annotated data, the generalization ability and the interpretability of AI methods remain as main challenges. In the future, it’s of great clinical significance to develop highly integrated intelligent radiological imaging systems with clear diagnosis evidence and general adaptation to complex clinical scenarios.