Improving Generalization of Deep Learning Models for Diagnostic Pathology by Increasing Variability in Training Data:...

Area of Science:

• Computational pathology
• Medical artificial intelligence

Background:

• Artificial intelligence (AI) shows promise in diagnostic pathology, with numerous deep learning studies on histopathological images.
• High variability in histopathological images can lead to overfitting and poor generalization in AI models, despite reported high accuracies.

Purpose of the Study:

• To illustrate the pitfalls of overfitting in AI models using osteosarcoma classification as an example.
• To demonstrate how incorporating model input variability enhances the performance of diagnostic AI.

Main Methods:

• Retrained a published osteosarcoma classification model using a publicly available dataset.
• Implemented a novel data partitioning strategy, isolating one patient's images for testing while training on others.
• Assessed the impact of training data variability by incrementally adding diverse osteosarcoma subtypes and benign tissues.

Main Results:

• The re-trained model exhibited a significant performance decline on the test set, confirming overfitting and lack of generalization.
• Stepwise addition of more subtypes to the training data progressively improved model performance.

Conclusions:

• Data preprocessing and collection strategies are crucial for managing high variability in histopathological images.
• These tactics help mitigate overfitting and develop more generalizable deep learning models for pathology.