Non-small cell lung cancer subtype classification based on cross-scale multi-instance learning.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Medical image analysis

Background:

• Non-small cell lung cancer (NSCLC) subtypes, including lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), pose diagnostic challenges impacting treatment.
• Accurate subtype classification is crucial for effective NSCLC treatment planning.

Purpose of the Study:

• To develop and validate a novel multi-instance learning (MIL) model for enhanced pathological image classification of NSCLC subtypes.
• To improve the accuracy and reliability of differentiating LUAD from LUSC using computational pathology.

Main Methods:

• A novel MIL model incorporating an additive attention mechanism and a category classifier was developed.
• A cross-scale focal region detection strategy was integrated to enhance feature sensitivity.
• The model was trained on the Cancer Genome Atlas (TCGA) dataset and validated on CPTAC TCIA and external hospital datasets.

Main Results:

• The model achieved 97.0% accuracy (ACC) and 0.978 area under the ROC curve (AUC) on the TCGA dataset, outperforming existing methods.
• Validation on external datasets showed robust performance with ACCs of 91.2% and 93.0%, and AUCs of 0.967 and 0.968.
• Ablation studies confirmed the significant contribution of each model component to performance.

Conclusions:

• The proposed MIL model demonstrates superior performance in classifying LUAD and LUSC subtypes.
• The model exhibits strong generalization capabilities across diverse datasets, indicating its potential for clinical application.
• This AI-driven approach offers a reliable tool for accurate NSCLC subtype diagnosis and improved patient management.