Pixel-Level Classification of Five Histologic Patterns of Lung Adenocarcinoma

Area of Science:

• Oncology
• Computational Pathology
• Digital Pathology

Background:

• Lung adenocarcinoma is the most common lung cancer subtype.
• Accurate histologic pattern classification is crucial for lung adenocarcinoma tumor grading.
• Current methods may lack the detailed pixel-level information needed for precise grading.

Purpose of the Study:

• To develop a novel Whole-Slide Image (WSI) analysis framework for pixel-level segmentation of lung adenocarcinoma histologic patterns.
• To enhance classification accuracy and robustness in digital pathology workflows.
• To provide pathologists with more detailed information for tumor grading.

Main Methods:

• Manual annotation of a dataset with 1000 patches (512x512) and 420 patches (1024x1024) across five histologic patterns.
• Development of a data stitching strategy for data augmentation.
• Implementation of a five-branch deep neural network framework for pattern segmentation.
• Evaluation using Dice Similarity Coefficient (DSC) and overall accuracy on test patches and WSIs.

Main Results:

• Data stitching improved DSC scores by 24.06% for the solid pattern on WSIs.
• The proposed WSI analysis framework outperformed existing networks (U-Net, LinkNet, FPN) with DSC scores up to 10.78% higher.
• The framework achieved an overall accuracy of 99.6% on four WSIs.

Conclusions:

• The developed WSI analysis framework accurately segments lung adenocarcinoma histologic patterns at the pixel level.
• The framework demonstrates superior accuracy and robustness compared to other networks.
• This technology has the potential to significantly assist pathologists in detailed lung adenocarcinoma tumor grading.