肺结节恶性风险预测模型研究进展

Advances in Predictive Models for Malignancy Risk in Pulmonary Nodules

摘要: 随着低剂量CT (LDCT)筛查的广泛应用，肺结节的检出率显著增加，这使得肺结节良恶性鉴别成为临床诊断中的关键挑战。传统的影像学评估方法依赖于医师的经验，存在较大主观性，且难以在复杂病例中准确区分结节的良恶性。近年来，数字诊断模型，尤其是影像组学(radiomics)和深度学习模型，逐渐成为肺结节管理的重要工具。这些模型通过提取和分析CT图像中的复杂特征，能够在更高维度上进行结节评估，显著提高了诊断的准确性。与此同时，多模态融合方法，结合临床数据、影像学特征和肿瘤标志物(如CEA、CYFRA 21-1)，为提高恶性预测能力提供了新的解决方案。同时，新型液体活检标志物及模型校准度优化逐渐成为提升模型临床适用性的研究热点。然而，尽管AI驱动的模型在精确性和智能化方面展现出巨大潜力，它们仍面临可解释性差、推广性有限以及数据偏移等问题。本文综述了肺结节风险评估模型的发展历程，探讨了传统模型与新兴AI方法的优势与局限，特别是在多模态融合与肿瘤标志物结合的潜力。最后，展望了未来数字诊断工具的发展方向，强调了提高可解释性、跨数据集验证和临床转化的必要性。

Abstract: With the widespread application of low-dose CT (LDCT) screening, the detection rate of pulmonary nodules has significantly increased, making the accurate differentiation between benign and malignant nodules a critical clinical challenge. Traditional imaging evaluation methods are highly subjective and often inadequate in distinguishing the malignancy of nodules, especially in complex cases. In recent years, digital diagnostic models, particularly imaging omics (radiomics) and deep learning models, have become essential tools in the management of pulmonary nodules. These models extract and analyze complex features from CT images, enabling a higher-dimensional assessment of nodules and significantly improving diagnostic accuracy. Furthermore, multi-modal integration methods combining clinical data, imaging features, and tumor markers (e.g., CEA, CYFRA 21-1) offer new solutions to enhance malignancy prediction. Emerging liquid biopsy biomarkers and model calibration optimization are also becoming important directions to improve real-world clinical applicability. Despite the substantial potential of AI-driven models in accuracy and intelligence, challenges such as poor interpretability, limited generalizability, and data bias still remain. This review systematically examines the evolution of pulmonary nodule risk assessment models, comparing traditional methods with emerging AI approaches, and highlights the potential and challenges of integrating tumor markers. Finally, we discuss future directions for the development of digital diagnostic tools, focusing on improving interpretability, cross-dataset validation, and clinical translation.