Attention Based Deep Multiple Instance Learning Approach for Lung Cancer Prediction using Histopathological Images

Area of Science:

• Pathology
• Computer Science
• Oncology

Background:

• Deep Convolutional Neural Networks (CNNs) are state-of-the-art for automated lung cancer diagnosis using histopathological images.
• Large datasets like The Cancer Genome Atlas have fueled advancements in CNN performance.
• Whole slide images often have weak annotations, limiting traditional supervised learning.

Purpose of the Study:

• To develop a Multiple Instance Learning (MIL) classifier for lung tissue type classification from whole slide images.
• To automate the detection of cancer in lung biopsy images.
• To enhance the interpretability and validation of automated diagnostic predictions.

Main Methods:

• Utilized Multiple Instance Learning (MIL) by treating whole slide images as bags of instances.
• Developed a bag/embedding-level classifier for lung tissue type determination.
• Employed a post-model interpretability algorithm for prediction validation and region highlighting.

Main Results:

• The proposed MIL model effectively classifies lung tissue types from whole slide images.
• Automated inspection of lung biopsies successfully identified the presence of cancer.
• Interpretability methods validated model predictions and pinpointed relevant regions of interest.

Conclusions:

• Multiple Instance Learning is a viable approach for weakly annotated histopathological data in lung cancer diagnosis.
• The developed classifier aids in automated lung cancer detection from whole slide images.
• Model interpretability is crucial for validating AI-driven diagnostic tools in pathology.