基于CT的慢性鼻窦炎分类的无分割三维深度学习

文献信息

中文摘要

背景： 慢性鼻窦炎（CRS）通常通过计算机断层扫描（CT）进行评估，但基于CT区分伴鼻息肉的慢性鼻窦炎（CRSwNP）与不伴鼻息肉的慢性鼻窦炎（CRSsNP）仍具有挑战性。方法：我们回顾性评估了一种无分割的三维卷积神经网络（3D CNN）用于三分类，使用经审核的776例成人CT检查队列：155例CRSwNP、222例CRSsNP和399例内分泌性眼眶病（EO）非CRS对照。体积被标准化为150×150×150体素。基线、增强和增强加L2配置仅在验证集宏F1上进行比较，之后才对选定的紧凑模型进行主要锁定测试评估；事后架构和敏感性分析随后在同一冻结测试集上进行，并被视为探索性分析。结果：在155例患者的留出测试集上，验证集选定的紧凑模型达到准确率0.748（95% CI 0.684-0.813）、平衡准确率0.683（0.602-0.763）、宏F1 0.691（0.607-0.770）和宏一对多AUC 0.882（0.841-0.919）。对EO对照的区分性能最强（一对多AUC 0.959），而直接的CRSsNP与CRSwNP区分较为有限（AUC 0.704；95% CI 0.574-0.827）。CRSwNP召回率为0.452，31例CRSwNP中有15例被分类为CRSsNP。因此，总体宏AUC主要受EO对照与CRS之间较强分离的影响，不应被解释为CRS表型区分 uniformly 强。在事后架构敏感性分析中，3D ResNet-10的观察测试宏F1更高，为0.742，但使用的参数约为紧凑模型的206倍。HiResCAM图在所有测试病例中技术上有效，尽管38/155张图（24.5%）的峰值位于强度衍生的患者包络之外，而该包络并非鼻窦分割掩膜。结论：这些发现为未来CRS亚型区分研究提供了研究阶段的原理验证；然而，潜在的人口学和采集相关捷径，以及缺乏外部验证，使得无法声称其具备临床就绪性。

英文摘要

Chronic rhinosinusitis (CRS) is routinely assessed with computed tomography (CT), but CT-based differentiation between CRS with nasal polyps (CRSwNP) and without nasal polyps (CRSsNP) remains challenging. We retrospectively evaluated a segmentation-free three-dimensional convolutional neural network (3D CNN) for three-class classification using an audited cohort of 776 adult CT examinations: 155 CRSwNP, 222 CRSsNP, and 399 endocrine orbitopathy (EO) non-CRS controls. The volumes were standardized to 150 × 150 × 150 voxels. Baseline, augmentation, and augmentation-plus-L2 configurations were compared exclusively by validation macro-F1 before the primary locked-test evaluation of the selected compact model; post hoc architecture and sensitivity analyses were subsequently performed on the same frozen test set and were considered exploratory. On the 155-patient held-out test set, the validation-selected compact model achieved an accuracy of 0.748 (95% CI 0.684-0.813), balanced accuracy of 0.683 (0.602-0.763), macro-F1 of 0.691 (0.607-0.770), and macro one-versus-rest AUC of 0.882 (0.841-0.919). Performance was strongest for EO-control discrimination (one-versus-rest AUC 0.959), whereas direct CRSsNP-versus-CRSwNP discrimination was more limited (AUC 0.704; 95% CI 0.574-0.827). The CRSwNP recall was 0.452, and 15 of 31 CRSwNP cases were classified as CRSsNP. Thus, the aggregate macro-AUC was influenced predominantly by the stronger separation of EO controls from CRS and should not be interpreted as uniformly strong CRS phenotype discrimination. In a post hoc architecture sensitivity analysis, a 3D ResNet-10 achieved a higher observed test macro-F1 of 0.742, but used approximately 206 times as many parameters as the compact model. The HiResCAM maps were technically valid in all test cases, although 38/155 maps (24.5%) peaked outside an intensity-derived patient envelope that was not a sinus segmentation mask. These findings provide a research-stage proof of concept for future studies on CRS subtype discrimination; however, potential demographic- and acquisition-related shortcuts, together with the absence of external validation, preclude any claim of clinical readiness.