Weakly supervised learning and interpretability for endometrial whole slide image diagnosis.

Area of Science:

• Digital Pathology
• Machine Learning
• Histopathology Imaging

Background:

• Fully supervised learning for histopathology diagnostics requires extensive expert annotation.
• Weakly supervised learning (WSL) offers a solution by using slide-level labels, reducing annotation burden.
• WSL has not been previously applied to endometrial whole slide images in iSyntax format.

Purpose of the Study:

• To apply a WSL algorithm to endometrial whole slide images for diagnostic tasks.
• To evaluate the performance and interpretability of the WSL model.
• To compare machine-learned insights with expert pathologist consensus.

Main Methods:

• Application of a weakly supervised learning algorithm to a real-world endometrial histopathology dataset.
• Utilized interpretability techniques: attention heatmapping, feature visualization, and end-to-end saliency mapping.
• Validated model performance using accuracy metrics on training, validation, and test sets.

Main Results:

• Achieved over 85% validation accuracy and over 87% test accuracy.
• Identified distinct learned morphologies through interpretability methods.
• Demonstrated a practical and robust application of WSL in digital pathology.

Conclusions:

• WSL is a viable and effective approach for endometrial whole slide image analysis.
• Interpretability methods are crucial for understanding and validating AI models in high-stakes medical applications.
• This study bridges the gap between machine learning capabilities and clinical expertise in histopathology.