Scholarlink

<p>Reliable lymph node metastasis (LNM) assessment underpins staging and treatment decisions in lung adenocarcinoma (LUAD), yet approximately 10% of patients remain pNx due to sampling limitations. Existing LNM prediction methods mainly rely on imaging or molecular data, while the potential of histopathological whole-slide images (WSIs) remains underexplored. We introduce CTMIL, a customized transformer-based multiple instance learning framework for predicting LNM directly from primary LUAD WSIs. Three CTMIL models were trained on 320 TCGA cases using patch-level features extracted from ResNet50, InceptionResNetV2, and UNI backbones, and internally validated on 80 held-out cases. External evaluation on two independent cohorts (XY cohort, 149 cases, and RM cohort, 164 cases) showed that CTMIL models consistently outperformed attention-based, TransMIL, and baseline approaches under identical inputs. The best-performing model, UNI_CTMIL, achieved AUROCs of 0.8640, 0.8216, and 0.8090 on the validation, XY, and RM cohorts, respectively. Interpretability analyses showed that UNI_CTMIL attends to metastatic hallmarks such as micropapillary and solid patterns, mucin, poor differentiation, and high tumor cell density. These findings highlight CTMIL’s promise for histology-based LNM prediction, potentially in pNx cases.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Customized transformer for lymph node metastasis prediction from lung adenocarcinoma histology in a multicentric study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月2日

查看全文
  • Huibo Zhang,
  • Tao Wang,
  • Junwei Feng,
  • Jie Wu,
  • Junju He,
  • Yang Liu,
  • Qibin Song,
  • Bin Xu

摘要

Reliable lymph node metastasis (LNM) assessment underpins staging and treatment decisions in lung adenocarcinoma (LUAD), yet approximately 10% of patients remain pNx due to sampling limitations. Existing LNM prediction methods mainly rely on imaging or molecular data, while the potential of histopathological whole-slide images (WSIs) remains underexplored. We introduce CTMIL, a customized transformer-based multiple instance learning framework for predicting LNM directly from primary LUAD WSIs. Three CTMIL models were trained on 320 TCGA cases using patch-level features extracted from ResNet50, InceptionResNetV2, and UNI backbones, and internally validated on 80 held-out cases. External evaluation on two independent cohorts (XY cohort, 149 cases, and RM cohort, 164 cases) showed that CTMIL models consistently outperformed attention-based, TransMIL, and baseline approaches under identical inputs. The best-performing model, UNI_CTMIL, achieved AUROCs of 0.8640, 0.8216, and 0.8090 on the validation, XY, and RM cohorts, respectively. Interpretability analyses showed that UNI_CTMIL attends to metastatic hallmarks such as micropapillary and solid patterns, mucin, poor differentiation, and high tumor cell density. These findings highlight CTMIL’s promise for histology-based LNM prediction, potentially in pNx cases.