超越全图学习:解剖分区深度学习模型用于上颌窦鼻腔疾病分类的优越性
多中心回顾性研究纳入2947例鼻窦CT,对150例手动分割13个解剖区域后,用nnU-Net v2自动分区并训练疾病特异网络,与全图分类器比较。解剖分区模型在外部测试集平均AUC为0.801,全图模型为0.587,73个诊断标签中42个AUC有统计学显著提升,平均绝对增幅0.214。
要点
- 解剖分区训练在13个解剖区域平均AUC为0.801,全图模型为0.587。
- 73个诊断标签中42个AUC有统计学显著提升,平均绝对增幅0.214。
- nnU-Net v2自动分区在13个区域平均Dice系数为0.739。
核心结论
- 解剖分区深度学习显著改善基于CT的鼻窦疾病诊断,较传统全图方法提供更可靠的病灶表征,并展现常规临床实施的较强潜力。
研究设计与样本
- 多中心回顾性研究纳入2019年10月至2022年5月的2947例CT检查。
- 随机选取150例患者进行手动分割,覆盖13个解剖区域,包括双侧鼻腔、上颌窦、前组筛窦、后组筛窦、蝶窦、额窦和鼻中隔。
- 先使用nnU-Net v2模型进行自动解剖分区,再在13个解剖分割的鼻窦亚区(每个扩展十个像素)上训练疾病特异网络,并与全图分类器比较。
- 使用外部测试数据评估敏感度、特异度和曲线下面积(AUC)。
诊断标准与关键数值
- 基于nnU-Net v2的解剖分区模型在13个解剖区域平均Dice系数为0.739。
- 外部测试队列中,解剖分区训练模型在13个解剖分区的平均AUC为0.801,全图训练模型在同一13个解剖分区的平均AUC为0.587。
- DeLong检验显示73个诊断标签中42个AUC有统计学显著改善,平均绝对增幅为0.214。
临床应用与阅片要点
- 作者认为解剖分区深度学习较传统全图方法提供更可靠的病灶表征,并展现常规临床实施的较强潜力。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVES: Whole-image deep learning models for CT diagnosis of nasal cavity and paranasal sinus diseases often underperform because they disregard anatomical heterogeneity. We hypothesized that anatomically partitioned deep learning would enhance diagnostic accuracy.
MATERIALS AND METHODS: A multicenter retrospective study enrolled 2947 CT examinations conducted from October 2019 to May 2022. From these, 150 patients were randomly selected for manual segmentation, covering 13 anatomical regions, including the bilateral nasal cavities, maxillary sinuses, anterior ethmoid sinuses, posterior ethmoid sinuses, sphenoid sinuses, frontal sinuses, and nasal septum. An nnU-Net v2 model was first used to perform automatic anatomical partitioning. Then, disease-specific networks were trained on the 13 anatomically segmented sinus subregions-each expanded by ten pixels-and compared with a whole-image classifier. External test data were used to evaluate sensitivity, specificity, and the area under the curve (AUC).
RESULTS: The anatomically partition model based on nnU-Net v2 achieved an average Dice coefficient of 0.739 across the 13 anatomical regions. For the external test cohort using the model trained with anatomical partitioning, the average AUC value for the 13 anatomical partitions was 0.801, whereas for the model trained on the whole image, the average AUC value for the same 13 anatomical partitions was 0.587. DeLong's test identified statistically significant improvements in AUC for 42 of the 73 diagnostic labels, with an average absolute increase of 0.214.
CONCLUSIONS: Anatomically partitioned deep learning markedly improves the CT-based diagnosis of sinonasal diseases, providing more reliable lesion characterization than conventional whole-image approaches and demonstrating strong potential for routine clinical implementation.
KEY POINTS: Question When whole-image deep learning models underperform across dozens of sinonasal pathologies, can an anatomically partitioned training strategy boost diagnostic performance on sinonasal CT? Findings Anatomically partitioned training decisively surpasses whole-image learning, delivering an average AUC improvement of 0.21 and establishing a potential framework for comprehensive sinonasal CT diagnosis. Clinical relevance This partitioned learning approach demonstrates a viable path for the clinical deployment of AI in sinonasal imaging, offering radiologists and rhinologists a reliable tool for multi-disease diagnosis that complements their expertise.
引用
- Li S, Hu XH, Luo S, et al. Beyond whole-image learning: anatomically partitioned deep learning models for superior sinonasal disease classification. Eur Radiol. 2026. PMID: 41832258. doi:10.1007/s00330-026-12436-4
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。