.png){kind=icon}
.png){kind=icon}
Return
Histopathology-Driven Multimodal Modeling for Lymph Node Metastasis Prediction and Survival Stratification in Lung Squamous Cell Carcinoma
.png){kind=icon}
PRE
AI
R
L
N
Q
X
T
J
M
K
X
H
DOI：10.1002/ima.70359.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To develop a multimodal computational pathology framework integrating handcrafted morphology and deep Vision Transformer (ViT) features to improve lymph node metastasis prediction and survival risk stratification in early-stage lung squamous cell carcinoma. Whole-slide images (WSIs) from 211 TCGA LUSC patients and an external cohort of 150 cases underwent standardized stain normalization, adaptive multiresolution patch extraction, and stringent artifact removal. Handcrafted morphological features were extracted using a customized CellProfiler pipeline, and deep embeddings were obtained from a fine-tuned ViT model. After ICC reliability filtering and correlation reduction, LASSO, mutual information (MI), and Boruta selection strategies were applied. A unified early-fusion vector was constructed and evaluated using SVM, XGBoost, TabNet, and TabTransformer classifiers. Survival analyses employed Cox-LASSO, Elastic Net, Random Survival Forests, and Gradient Boosting Survival models, with performance assessed using AUC, C-index, IBS, and time-dependent AUC. Deep ViT-derived features outperformed handcrafted morphology across all metrics (external AUC 0.82 vs. 0.71), and fusion significantly enhanced predictive performance, achieving test and external AUCs up to 0.94 and 0.92, respectively. The LASSO-TabTransformer fusion pipeline provided the strongest metastasis discrimination, with recall approaching 87%. Survival modeling demonstrated similarly substantial improvements: fusion-based Gradient Boosting Survival achieved C-indices of 0.91 (test) and 0.88 (external), with 1-year AUCs of 0.92 and 0.89. FDR-validated imaging covariates remained independently prognostic, and Kaplan–Meier curves confirmed robust stratification of LNM+ versus LNM− groups. Multimodal fusion of deep and handcrafted features produces highly accurate, generalizable models for LNM prediction and survival stratification in LUSC, offering a powerful tool for preoperative decision support.
Keywords:
lung squamous cell carcinoma
lymph node metastasis prediction
machine learning
multimodal feature fusion
survival modeling
vision transformer embeddings
whole-slide imaging
Journal
IF:
2.5
Papers:
2.1K
Citations:
2.3K
Organization
Cited Papers