Lung cancer subtype classification using histopathological images based on weakly supervised multi-instance learning.

Area of Science:

• Digital pathology
• Computational oncology
• Artificial intelligence in medicine

Background:

• Accurate non-small-cell lung cancer (NSCLC) subtype classification is crucial for guiding clinical diagnosis and treatment.
• Whole slide images (WSIs) present challenges for automated analysis due to their large size and lack of distinct morphological features, often necessitating manual region of interest (ROI) delineation.

Purpose of the Study:

• To develop a weakly supervised framework for accurate NSCLC subtype classification from WSIs, eliminating the need for manual pixel-level annotation.
• To propose a two-stage approach for ROI localization and subsequent subtype classification.

Main Methods:

• A multi-resolution expectation-maximization convolutional neural network (MR-EM-CNN) was developed for ROI localization, utilizing the EM algorithm to identify discriminative patches from WSI-wise labels.
• A hierarchical attention multi-scale network (HMS) was designed for subtype classification, capturing multi-scale features and enabling hierarchical feature interaction.

Main Results:

• The proposed framework achieved an Area Under the Curve (AUC) of 0.9602 for ROI localization.
• The subtype classification achieved an AUC of 0.9671 on the Cancer Genome Atlas dataset comprising 1002 patients.

Conclusions:

• The developed weakly supervised framework demonstrates superior performance in NSCLC subtype classification compared to existing methods.
• The proposed approach offers a scalable solution that can be extended to other WSI-based classification tasks.