A novel transfer-learning based physician-level general and subtype classifier for non-small cell lung cancer

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Accurate histological subtyping of lung adenocarcinoma is crucial for patient prognosis and treatment decisions.
• Pathologist interobserver concordance for lung adenocarcinoma subtypes can be limited, highlighting the need for objective diagnostic tools.
• Computer-aided diagnosis (CADx) systems offer potential to expedite diagnosis and improve consistency.

Purpose of the Study:

• To develop and validate a novel computer-aided diagnosis method for accurate lung adenocarcinoma subtype classification.
• To reduce the reliance on extensive manual annotations by pathologists through integrated learning and transfer learning.
• To enhance the efficiency and accuracy of lung adenocarcinoma histological pattern identification.

Main Methods:

• A hybrid approach combining weak supervised learning and integrated learning with transfer learning was employed.
• Two large datasets, The Cancer Genome Atlas (TCGA) and Clinical Proteomic Tumor Analysis Consortium (CPTAC), were utilized.
• Whole-slide images (WSI) were processed to classify adenocarcinoma and identify subtypes using weak classifiers and integrated learning.

Main Results:

• The proposed model achieved high Area Under the Curve (AUC) values across various subtypes: Acinar (0.86), LPA (0.91), Micropapillary (0.82), Papillary (0.77), Solid (0.96), and Normal (0.98).
• Validation on independent datasets from CPTAC and a private hospital demonstrated the model's robust performance.
• The method effectively reduced the requirement for manual pathological annotations while maintaining diagnostic accuracy.

Conclusions:

• The developed computer-aided diagnosis system demonstrates significant potential for accurate and efficient lung adenocarcinoma subtyping.
• This approach, leveraging integrated learning and transfer learning, can assist pathologists and improve diagnostic workflows.
• The findings suggest a promising direction for AI-driven pathological analysis in oncology, aiding in personalized patient care.