评估可解释的人工智能技术在多种成像方式下解释乳腺癌诊断的有效性

Evaluating the Effectiveness of Explainable AI Techniques in Interpreting Breast Cancer Diagnoses Across Multiple Imaging Modalities

摘要: 乳腺癌持续位居全球女性癌症发病与致死的主要原因之列。早期且精确的诊断对于优化患者预后具有举足轻重的地位。乳房X线摄影、超声检查及磁共振成像(Magnetic Resonance Imaging, MRI)等影像学技术在乳腺癌的诊断中扮演着至关重要的角色。然而，这些技术手段面临着准确性波动、操作员依赖性显著及结果阐释困难等多重挑战。在此背景下，人工智能(Artificial Intelligence, AI)，尤其是可解释人工智能(Explainable Artificial Intelligence, XAI)的融入，已成为提升诊断精确度及增强信任度的革命性途径。本综述聚焦于XAI技术在乳腺癌诊断领域内，于不同成像模式中的应用效果比较。深入探讨了核心的XAI方法，诸如Shapley加性解释(SHAP)、局部可解释模型无关解释(LIME)以及基于梯度的类激活映射(Grad-CAM)，着重阐述了它们在增进模型可解释性及提升临床实用性方面的具体成效。综述不仅分析了XAI技术在乳房X线摄影、超声及MRI应用中的优势与局限，还特别强调了其在提高AI辅助预测透明度方面的贡献。此外，本文亦评估了XAI在应对假阳性、假阴性问题以及多模态成像数据整合挑战中的效能。该评论的核心价值在于，它全面剖析了XAI在缩小AI技术进展与临床实际应用之间鸿沟的潜力。通过提升透明度，XAI技术能够增强临床医生对AI的信任度，促进其更顺畅地融入诊断工作流程，从而助力个性化医疗实践的推进及患者治疗成效的改善。综上所述，尽管XAI在提升AI模型可解释性与准确性方面取得了显著进展，但在计算复杂度控制、普遍适用性拓展及临床接纳度提升等方面仍面临诸多挑战。未来研究应着重于优化XAI方法、促进跨学科间的深度合作，并开发标准化的框架体系，以确保XAI技术能在多样化的临床环境中实现可扩展性与可靠性的双重提升。

Abstract: Breast cancer remains one of the leading causes of cancer incidence and mortality among women worldwide. Early and accurate diagnosis plays a pivotal role in optimizing patient prognosis. Imaging techniques such as mammography, ultrasound, and magnetic resonance imaging (MRI) play crucial roles in the diagnosis of breast cancer. However, these techniques face multiple challenges, including accuracy fluctuations, significant operator dependency, and difficulties in result interpretation. In this context, the integration of Artificial Intelligence (AI), especially Explainable Artificial Intelligence (XAI), has become a revolutionary approach to improving diagnostic accuracy and enhancing trust. This review focuses on the comparative application of XAI technologies across different imaging modalities in breast cancer diagnosis. It delves into core XAI methods such as Shapley Additive Explanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Gradient-weighted Class Activation Mapping (Grad-CAM), with an emphasis on their effectiveness in enhancing model interpretability and improving clinical utility. The review analyzes not only the advantages and limitations of XAI in mammography, ultrasound, and MRI applications but also highlights its contribution to increasing the transparency of AI-assisted predictions. Additionally, the review evaluates the performance of XAI in addressing issues related to false positives, false negatives, and the challenges of multimodal imaging data integration. The core value of this review lies in its comprehensive analysis of the potential of XAI in bridging the gap between advancements in AI technology and clinical application. By enhancing transparency, XAI can boost clinicians’ trust in AI, facilitating its smoother integration into diagnostic workflows, thereby promoting personalized medical practices and improving patient treatment outcomes. In conclusion, despite significant progress made by XAI in improving AI model interpretability and accuracy, challenges remain in terms of computational complexity, general applicability, and clinical acceptance. Future research should focus on optimizing XAI methods, fostering interdisciplinary collaboration, and developing standardized frameworks to ensure the scalability and reliability of XAI technologies in diverse clinical environments.