OKEN: A Supervised Evolutionary Optimizable Dimensionality Reduction Framework for Whole Slide Image Classification

Area of Science:

• Digital pathology and computational oncology
• Artificial intelligence in medical image analysis

Background:

• Accurate lung cancer subtype classification is crucial for clinical treatment.
• Histopathology image analysis faces challenges, often requiring immunohistochemistry.
• Digital pathology and AI offer automated solutions for tissue slide analysis.

Purpose of the Study:

• To develop an AI framework for Whole Slide Image (WSI) classification of lung cancer.
• To utilize an evolutionary algorithm for feature encoding into an adjustable latent space.
• To improve computational efficiency for WSI classification and segmentation.

Main Methods:

• Development of a WSI classification framework with an optimizable kernel.
• Feature encoding from WSI patches into a latent space using an evolutionary algorithm.
• Comparison with the state-of-the-art Vim4Path model on internal and external lung cancer WSI datasets.

Main Results:

• The proposed framework outperformed Vim-S16 in accuracy and F1 score on an internal dataset at ×2.5 and ×10 magnification.
• Highest accuracy (0.833) and F1 score (0.721) achieved at ×2.5 magnification on the internal test set.
• On the external test set, Vim-S16 at ×10 had the highest accuracy (0.732), while OKEN-DenseNet121 at ×2.5 had the best F1 score (0.772).

Conclusions:

• The developed AI framework shows promise for accurate and efficient lung cancer WSI classification.
• The evolutionary algorithm-based feature encoding offers a competitive approach compared to existing models.
• Future work should focus on dynamically tuning the output dimensions of the evolutionary algorithm.