Abstract
Lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), the two major subtypes of non-small cell lung cancer (NSCLC), present significant diagnostic challenges with direct implications for treatment planning. In this study, we propose a novel multi-instance learning (MIL) pathological image classification model that incorporates an additive attention mechanism and a new category classifier to enhance subtype discrimination. The model further integrates a cross-scale focal region detection strategy to improve sensitivity to key histological features. Trained on the Cancer Genome Atlas (TCGA) dataset, our model achieved a subtype classification accuracy (ACC) of 97.0% and an area under the ROC curve (AUC) of 0.978, outperforming state-of-the-art methods including ABMIL, CLAM, DS-MIL, DTFD-MIL, FR-MIL, and WIKG-MIL across multiple evaluation metrics. Ablation studies validate the contribution of each module to overall performance improvement. Generalization experiments conducted on the Clinical Proteomic Tumor Analysis Consortium (CPTAC) Cancer Imaging Archive (TCIA) Lung dataset and an external dataset from Yantai Yuhuangding Hospital demonstrate the robustness of our model, achieving ACCs of 91.2% and 93.0%, and AUCs of 0.967 and 0.968, respectively. These results underscore the model's strong generalization ability and its potential as a reliable tool for accurate NSCLC subtype classification across diverse clinical scenarios.