Related Experiment Videos
Integration of Histologic Pattern Classifiers and Graph Convolutional Networks for Prognostic Prediction in Lung
Yi-Chen Yeh1,2, Wei-Hsiang Yu3, Min-Shu Hsieh4,5,6
1Department of Pathology and Laboratory Medicine, Taipei Veterans General Hospital, Taipei, Taiwan.
Introduction:
Histologic pattern classification is crucial for diagnosing and prognosticating lung adenocarcinoma, yet its clinical application remains limited by interobserver variability and difficulties in quantifying patterns across whole-slide images (WSIs).
Methods:
We developed a deep learning-based histologic pattern classifier trained with coarse annotations and implemented through a whole-slide approach. Classification performance was evaluated across five major growth patterns, followed by consensus analysis with expert pathologists. To predict prognosis, we integrated the classifier into a patch-level graph convolutional network (patch-GCN) that aggregated local and global morphologic features across WSIs. Model generalizability was further assessed in two external cohorts.
Results:
The classifier achieved a mean area under the curve of 0.991 and 90.5% expert model agreement. Consensus analysis revealed strong concordance on typical morphologies but underscored challenges for both experts and the model in regions with atypical histology. The patch-GCN achieved the best prognostic performance (concordance index: 0.789), outperforming models based solely on expert-assessed or model-predicted histologic pattern percentages. Notably, the GCN identified prognostically adverse morphologies, including complex glandular and morule-like patterns, which were not explicitly annotated during training. External validation demonstrated robust generalizability, with concordance indices of 0.654 in the TCGA-LUAD cohort and 0.868 in the NTUH cohort. Model performance depended strongly on pathology-specific feature embeddings.
Conclusion:
This study demonstrates the feasibility of a WSI-based histologic pattern classifier trained with coarse annotations and establishes a novel GCN-based prognostic framework that enhances predictive accuracy, uncovers clinically relevant morphologies beyond expert labels, and supports granular risk stratification in lung adenocarcinoma.