Lung cancer is the second most common type of cancer worldwide, making up about 20% of all cancer deaths with less than 10% 5-year survival rate for the very late stage. The recent guidelines for the most common non-small-cell lung cancer (NSCLC) type recommend performing staging based on the 8th edition of TNM classification, where the mediastinal lymph node involvement plays a key role. However, most of the non-invasive methods have a very limited level of sensitivity and are relatively accurate, but invasive methods can be contradicted for some patients. Current advances in Deep Learning show great potential in solving such problems. Still, most of these works focus on the algorithmic side of the problem, not the clinical relevance. Moreover, none of them addressed individual lymph node malignancy classification problem, restricting the indirect analysis of the whole study, and limiting the interpretability of the result without giving an option for cliniciansto validate the result. This work mitigates these gaps, proposing a multi-step algorithm for each visible mediastinal lymph node segmentation and assessing the probability of its involvement in themetastatic process, using the results of histological verification on training. The developed pipelineshows 0.74 ± 0.01 average Recall with 0.53 ± 0.26 object Dice Score for the clinically relevant lymph nodes segmentation task and 0.73 ROC AUC for patient’s N-stage prediction, outperformingtraditional size-based criteria.

中文翻译：

---

使用卷积神经网络对计算机断层扫描进行自动淋巴结分割和组织学状态分类

肺癌是全球第二常见的癌症，约占所有癌症死亡的 20%，晚期的 5 年生存率不到 10%。最近针对最常见非小细胞肺癌 (NSCLC) 类型的指南建议根据第 8 版 TNM 分类进行分期，其中纵隔淋巴结受累起着关键作用。然而，大多数非侵入性方法的灵敏度非常有限并且相对准确，但侵入性方法对于某些患者可能会产生矛盾。当前深度学习的进展显示出解决此类问题的巨大潜力。尽管如此，这些工作大多数都集中在问题的算法方面，而不是临床相关性。此外，它们都没有解决个体淋巴结恶性肿瘤的分类问题，限制了整个研究的间接分析，并限制了结果的可解释性，而没有为临床医生提供验证结果的选择。这项工作弥补了这些差距，为每个可见的纵隔淋巴结分割提出了一种多步骤算法，并使用训练的组织学验证结果评估其参与转移过程的概率。开发的流程显示，临床相关淋巴结分割任务的平均召回率为 0.74 ± 0.01，对象 Dice 评分为 0.53 ± 0.26，患者 N 期预测的 ROC AUC 为 0.73，优于传统的基于大小的标准。