Transformaer-based model for lung adenocarcinoma subtypes.

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Lung cancer has the highest morbidity and mortality rates globally.
• Histological subtypes are critical for lung cancer diagnosis, prognosis, and treatment response prediction.
• Existing methods fail to capture nuanced characteristics of lung adenocarcinoma (LAD) tissue subtypes.

Purpose of the Study:

• To pioneer the classification of LAD into five distinct histological subtypes: acinar, lepidic, micropapillary, papillary, and solid.
• To develop and validate a novel deep learning model, HybridNet, for improved LAD subtype classification.

Main Methods:

• HybridNet employs a dual-stream architecture, integrating a Transformer for global representations and a Convolutional Neural Network (CNN) for local features.
• Self-attention mechanisms in the Transformer capture rich contextual information.
• Feature maps from both streams are interactively fused to enhance discriminative power for classification.

Main Results:

• HybridNet achieved 95.12% overall classification accuracy on a private LAD dataset.
• Individual subtype accuracies were: acinar (94.5%), lepidic (97.1%), micropapillary (94%), papillary (91%), and solid (99%).
• The model demonstrated strong performance on the public BreakHis dataset, achieving the best results in accuracy (92.40%), recall (90.63%), and F1-score (91.43%).

Conclusions:

• Classifying LAD into five subtypes aids pathologists in treatment selection, tumor mutation burden (TMB) prediction, and immune checkpoint protein analysis.
• HybridNet effectively fuses CNN and Transformer features, significantly improving subtype classification accuracy.
• The model exhibits satisfactory generalization ability on public datasets, indicating its potential clinical utility.