Page 37 - 《中国医疗器械杂志》2026年第1期
P. 37
Chinese Journal of Medical Instrumentation 2026年 第50卷 第1期
医 学 人 工 智 能
PEAIM保持了优于单器官模型的稳定性能,这验 [4] TESSLER F N, MIDDLETON W D, GRANT E G, et al.
证了其面对不同医疗中心数据时的鲁棒性。在迁移 ACR Thyroid Imaging, Reporting and Data System (TI-
RADS): white paper of the ACR TI-RADS Committee[J].
性方面,多任务学习框架和模型的分层设计(共享
J Am Coll Radiol, 2017, 14(5): 587-595.
层+特异层)使知识能够在两类结节任务间有效迁
[5] D'ORSI C J, SICKLES E A, MENDELSON E B, et al.
[34]
移 。实验中可观察到,随着样本量减少,PEAIM ACR BI-RADS® Atlas, Breast Imaging Reporting and
相比单器官模型的优势逐渐增大,证实了PEAIM Data System[M]. Reston, VA: American College of
在数据稀疏情境下的良好迁移学习能力。特别是在 Radiology, 2013.
[6] CIBAS E S, ALI S Z. The 2017 Bethesda system for
小样本学习实验(使用10%训练数据)中,PEAIM
reporting thyroid cytopathology[J]. Thyroid, 2017,
仍保持AUC>0.85的性能,而单器官模型性能则降
27(11): 1341-1346.
至AUC<0.80。这种强大的迁移学习能力对于罕见 [7] HOANG J K, MIDDLETON W D, FARJAT A E, et al.
[35]
亚型结节和数据有限的临床场景具有重要价值 。 Interobserver variability of sonographic features used in
本模型也存在技术局限:①计算复杂度较高, the American College of Radiology Thyroid Imaging
Transformer架构需要较大的计算资源和内存,可能 Reporting and Data System[J]. AJR Am J Roentgenol,
2018, 211(1): 162-167.
限制其在资源设备上的应用;②对特征分布的依赖
[8] ELEZABY M, LI G, BHARGAVAN-CHATFIELD M, et
性,当新数据分布与训练集显著偏离时(如不同种 al. ACR BI-RADS assessment category 4 subdivisions in
族人群或不同超声设备),模型性能可能下降,需 diagnostic mammography: utilization and outcomes in the
要额外的域适应技术;③多任务学习中存在任务冲 National Mammography Database[J]. Radiology, 2018,
突问题:部分特征对不同器官结节的预测任务影响 287(2): 416-422.
[9] GAO Y, GOLDBERG J E, YOUNG T K, et al. Breast
方向相反(如年龄因素在甲状腺结节中与恶性风险
cancer screening with digital breast tomosynthesis and
负相关,在乳腺结节中则正相关),这种冲突可能 digital mammography in dense and nondense breasts[J].
导致模型共享表示层的特征学习不够优化,因此需 JAMA, 2020, 323(8): 746-756.
要设计更精细的特征分离机制以区分跨器官共性特 [10] ZHU Y L, SONG Y T, XU G H, et al. Causes of
征与器官特异性特征;④可解释性挑战,虽然使用 misdiagnoses by thyroid fine-needle aspiration cytology
(FNAC): our experience and a systematic review[J].
SHAP提供了特征重要性解释,但Transformer的注
Diagn Pathol, 2020, 15(1): 1
意力机制和深层表示学习仍有“黑盒”特性,临床 [11] LI X, ZHANG S, ZHANG Q, et al. Diagnosis of thyroid
医师难以完全理解具体决策路径;⑤数据质量敏感 cancer using deep convolutional neural network models
性,模型对标签质量和特征规范化程度较为敏感, applied to sonographic images: a retrospective,
这可能为其多中心实施带来标准化挑战。未来工作 multicohort, diagnostic study[J]. Lancet Oncol, 2019,
20(2): 193-201.
将致力于开发更轻量级的模型架构、改进域适应技
[12] WANG L, YANG S, YANG S, et al. Automatic thyroid
术、设计更精细的特征共享机制、增强模型可解释 nodule recognition and diagnosis in ultrasound imaging
性,以及建立更严格的数据质量控制流程,从工程 with the YOLOv2 neural network[J]. World J Surg
角度进一步提升模型的临床实用性。 Oncol, 2019, 17(1): 12.
[13] WU G G, ZHOU L Q, XU J W, et al. Artificial
intelligence in thyroid ultrasound diagnosis[J]. Med Phys,
参考文献
[1] HAUGEN B R, ALEXANDER E K, BIBLE K C, et al. 2022, 49(9): 5974-5996.
2015 American Thyroid Association management [14] WANG P, XIAO X, GLISSEN BROWN J R, et al.
guidelines for adult patients with thyroid nodules and Development and validation of a deep-learning algorithm
differentiated thyroid cancer: the American Thyroid for the detection of polyps during colonoscopy[J]. N Engl
Association Guidelines Task Force on Thyroid Nodules J Med, 2020, 382(22): 2107-2118.
and Differentiated Thyroid Cancer[J]. Thyroid, 2016, [15] MCKINNEY S M, SIENIEK M, GODBOLE V, et al.
26(1): 1-133. International evaluation of an AI system for breast cancer
[2] BRITT K L, CUZICK J, PHILLIPS K A. Key steps for screening[J]. Nature, 2020, 577(7788): 89-94.
effective breast cancer prevention[J]. Nat Rev Cancer, [16] YALA A, MIKHAEL P G, STRAND F, et al. Toward
2020, 20(8): 417-436. robust mammography-based models for breast cancer
[3] SIEGEL R L, MILLER K D, FUCHS H E, et al. Cancer risk[J]. Sci Transl Med, 2021, 13(578): eaba4373.
statistics, 2022[J]. CA Cancer J Clin, 2022, 72(1): 7-33. [17] ZHANG L, WU Y, LIU F, et al. Characteristics and
33

