Classifying non-small cell lung cancer types and transcriptomic subtypes using convolutional neural networks.

Area of Science:

• Computational pathology
• Oncology
• Bioinformatics

Background:

• Non-small cell lung cancer (NSCLC) is a major cause of cancer mortality globally.
• Histopathological evaluation is crucial for NSCLC diagnosis, but morphological patterns of molecular subtypes remain understudied.
• Objective classification of NSCLC subtypes is needed to improve diagnosis and treatment.

Purpose of the Study:

• To develop a quantitative histopathology analytic framework for objective identification of NSCLC types and gene expression subtypes.
• To bridge the gap between morphological patterns and molecular subtypes in NSCLC.
• To create an automated system for NSCLC subtyping.

Main Methods:

• Utilized whole-slide histopathology images from lung adenocarcinoma and squamous cell carcinoma patient cohorts in The Cancer Genome Atlas.
• Developed and validated convolutional neural networks (CNNs) for image classification and tumor region identification.
• Assessed CNN performance using area under the receiver-operating characteristic curves (AUCs) and validated in an independent cohort.

Main Results:

• CNN models accurately identified tumor regions (AUCs > 0.935) and recapitulated expert pathologist diagnoses (AUCs > 0.877).
• Quantitative morphology features derived from CNNs successfully identified major transcriptomic subtypes in both lung adenocarcinoma and squamous cell carcinoma (P < .01).
• Validation in an independent cohort demonstrated robust performance (AUCs = 0.726–0.864).

Conclusions:

• This study presents the first fully automated machine learning method for classifying NSCLC transcriptomic subtypes.
• The developed framework objectively identifies novel, clinically relevant histopathology patterns without prior pathology knowledge.
• The methodology is generalizable to other cancer types and diseases.