Effects of Image Quantity and Image Source Variation on Machine Learning Histology Differential Diagnosis Models.

Area of Science:

• Medical Education
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Histology is fundamental to medical training, requiring diagnostic competency.
• Machine learning (ML) offers potential for developing advanced histology learning tools.
• Convolutional neural networks (CNNs) are a type of ML model with potential for image classification.

Purpose of the Study:

• To develop a generalizable multi-classification model using a CNN for classifying human tissue images.
• To assess the model's ability to provide differential diagnoses for histological entities.
• To explore the impact of training data characteristics on model generalizability.

Main Methods:

• Acquired three institutional and one generalizability test datasets with 38 tissue categories.
• Trained CNN models using varying data quantities and sources (single vs. multiple institutions).
• Tested model performance on internal validation, external institutional, and Google image search data using accuracy, sensitivity, specificity, and F1-score.

Main Results:

• Model generalizability depends on training data source diversity and quantity.
• Models trained on single-institution data showed poor external performance.
• Increasing data source diversity improved generalizability, even with fewer images.
• Optimal generalizability was achieved with diverse, high-quantity training datasets (2280 images from multiple sources).

Conclusions:

• Data diversity is crucial for developing generalizable ML models in histology.
• Optimal ML model performance requires incorporating both data diversity and quantity.
• This approach can enhance histology learning platforms and diagnostic support.