Which data subset should be augmented for deep learning? a simulation study using urothelial cell carcinoma...

Area of Science:

• Digital pathology
• Machine learning
• Deep learning

Background:

• Deep learning for digital histopathology is limited by scarce annotated data.
• Data augmentation methods are not standardized, impacting model development.
• This study systematically compares 11 augmentation strategies.

Purpose of the Study:

• To explore the impact of data augmentation timing and dataset subset application on deep learning model performance in digital histopathology.
• To identify optimal data augmentation strategies for improving model accuracy and reliability.

Main Methods:

• Compared 11 data augmentation strategies, varying timing and dataset subsets (training, validation, test).
• Utilized hematoxylin-and-eosin-stained urinary bladder slides, classifying images into inflammation, urothelial cell carcinoma, or invalid categories.
• Fine-tuned four pre-trained convolutional neural networks (Inception-v3, ResNet-101, GoogLeNet, SqueezeNet) for binary classification.
• Evaluated model performance using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve.

Main Results:

• Augmentation applied after test-set separation but before training/validation split yielded the best testing performance, though with optimistic validation accuracy due to information leakage.
• Augmentation before test-set separation also led to optimistic results.
• Augmenting the test set provided more accurate evaluation metrics with reduced uncertainty.
• Inception-v3 demonstrated the best overall testing performance among the evaluated models.

Conclusions:

• For digital histopathology, data augmentation should encompass the test set (post-allocation) and the combined training/validation set (pre-split).
• This approach balances performance optimization with reliable evaluation.
• Further research is needed to generalize these findings across different datasets and tasks.