Preparing pathological data to develop an artificial intelligence model in the nonclinical study

Area of Science:

• Digital pathology
• Computational pathology
• Histopathology image analysis

Background:

• Artificial intelligence (AI) is increasingly used for analyzing digitized histological slides.
• Hematoxylin and eosin (H&E) staining is standard, but variations in color tone and magnification can affect AI model performance.
• Whole slide images (WSIs) offer comprehensive tissue visualization.

Purpose of the Study:

• To investigate the impact of staining color tone and magnification variations on AI model predictions using H&E stained WSIs.
• To compare the performance of AI models trained on single versus mixed datasets of varying image characteristics.
• To optimize AI model training for consistent and accurate pathological lesion detection.

Main Methods:

• Prepared three datasets (N20, B20, B10) of liver tissue WSIs with fibrosis, varying color tones and magnifications.
• Trained five Mask R-CNN models using single or mixed combinations of these datasets.
• Evaluated model performance on a separate test dataset comprising all three variations.

Main Results:

• Models trained on mixed datasets (B20/N20, B10/B20) demonstrated superior performance compared to models trained on single datasets.
• The enhanced performance of mixed models was confirmed by actual prediction results on test images.
• Variations in staining color tone and magnification were key factors influencing AI prediction outcomes.

Conclusions:

• Training AI algorithms with diverse staining color tones and multi-scaled image datasets leads to more robust and consistent performance.
• Mixed datasets are crucial for developing reliable AI tools for pathological lesion prediction in digital pathology.
• Future AI development should incorporate dataset variability for improved diagnostic accuracy.