Transformer-Based Weakly Supervised Learning for Whole Slide Lung Cancer Image Classification.

Area of Science:

• Computational pathology
• Artificial intelligence in oncology
• Deep learning for medical imaging

Background:

• Whole-slide images (WSIs) are crucial for lung cancer diagnosis.
• Manual pixel-wise annotation of WSIs is labor-intensive and prone to variability.
• Weakly supervised learning offers a promising alternative to overcome annotation limitations.

Purpose of the Study:

• To develop an effective weakly supervised learning framework for histopathological lung cancer diagnosis.
• To address the challenges of large-scale WSI analysis and annotation variability.
• To improve the accuracy and efficiency of lung cancer subtype classification.

Main Methods:

• Proposed a two-stage transformer-based weakly supervised learning framework: Simple Shuffle-Remix Vision Transformer (SSRViT).
• Introduced Shuffle-Remix Vision Transformer (SRViT) for discriminative token retrieval and feature extraction.
• Utilized a simple transformer-based classifier (SViT) for slide-level prediction using aggregated WSI features.

Main Results:

• SSRViT achieved high performance in discriminating between adenocarcinoma, pulmonary sclerosing pneumocytoma, and normal lung tissue.
• Demonstrated significant improvement over state-of-the-art methods.
• Attained an accuracy of 96.9% and an Area Under the Curve (AUC) of 99.6%.

Conclusions:

• The proposed SSRViT framework effectively leverages weak labels for histopathological lung cancer diagnosis.
• SSRViT offers a robust and accurate solution for classifying lung cancer subtypes from WSIs.
• This approach has the potential to streamline diagnostic workflows and reduce inter-observer variability.