Multiview 2.5D Deep Learning Outperforms 2D and 3D Models for Preoperative Prediction of Visceral Pleural Invasion in...

This study evaluated the predictive performance of 2 novel 2.5-dimensional (2.5D) deep learning (DL) models for visceral pleural invasion (VPI) in clinical stage IA lung adenocarcinoma, comparing them with traditional 2D and 3D models.

A multicenter retrospective analysis included 804 patients from 2 Chinese hospitals with pathologically confirmed stage IA lung adenocarcinoma. The cohort was divided into training (n=360), internal validation (n=155), and external test sets (n=289). Two 2.5D models were developed: a multiview model integrating the largest tumor sections from coronal, sagittal, and axial planes, and a context model incorporating the largest axial slice with adjacent slices. Model performance was assessed using pathological diagnosis as the reference standard, with discriminative abilities evaluated via area under the curve (AUC), accuracy, and predictive values.

The 2.5D multiview model achieved an AUC of 0.73 in external validation, outperforming 2D (axial: 0.66, coronal: 0.64, sagittal: 0.63), 2.5D context (0.67), and 3D models (0.66). It demonstrated 67% accuracy, 48% positive predictive value (PPV), and 85% negative predictive value (NPV). Grad-CAM visualization highlighted tumor-pleura contact zones and peritumoral regions as critical predictors of VPI.

The 2.5D multiview DL model enhances preoperative VPI prediction in stage IA lung adenocarcinoma, offering superior accuracy and interpretable insights to guide surgical decisions.

This study is the first to validate 2.5D DL models, particularly the multiview approach, for VPI prediction, demonstrating improved performance over 2D/3D models while revealing spatially relevant biomarkers through interpretable visualization.