CPF-Net: Cross-modal CT and Pathology Guided Feature Learning for CT-based Lung Cancer Subtype Classification

Abstract

Accurate classification of lung cancer subtypes from CT images remains challenging due to the subtle radiological differences between adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC). We propose CPF-Net, a deep learning framework that integrates CT and pathological information through a Linear Spatial Reduction Attention (LSRA) module. The framework processes whole slide images using a modified CTransPath architecture for pathological feature extraction and combines these features with CT imaging characteristics during training. While both CT and pathological data are used in training, only CT images are required for inference. Experiments on a dataset of 892 cases from The Cancer Genome Atlas (TCGA) show that CPF-Net achieves 87.89% accuracy, 93.23% AUC, and 86.92% F1-score, outperforming existing methods by margins of 4.44%, 3.67%, and 4.14% respectively. Ablation studies demonstrate the effectiveness of both the LSRA module and the cross-modal learning strategy in improving classification performance.