Sample Size Estimation of Data Augmentation Strategies in Colorectal Cancer Histology via Transfer Learning

Microsatellite instability (MSI) status classification based on colorectal cancer histology using deep learning can be challenging when training data are limited. For exploration, this study investigates traditional data augmentation strategies, quantifying their impact on sample size requirements and classification accuracy in transfer learning for colorectal cancer histology images. In this study, various conventional augmentation strategies, including brightness adjustment, rotation, noise addition, and mixed augmentation, were analysed based on sample size estimation using curve fitting with the inverse power law. To generate the learning curves, an external K-fold cross-validation loop was implemented with each different sample size (ranging from 10 to 200 image patches). Additionally, three popular deep neural networks (DenseNet121, InceptionV3, and MobileNetV2) were utilised with three model transfer methods: using the original network (‘ori’), fine-tuning the last layers (‘ft’), and adding extra convolutional layers (‘add’). After curve fitting, summary plots and their corresponding difference plots were created to assess data augmentation strategies as well. Based on the exploration of this study, it has been found that simple model transfer methods (e.g., ‘ori’ and ‘ft1’) remain favourable for data-limited applications, even with conventional data augmentation. In general, not all combinations of model transfer methods and data augmentation strategies guarantee an improvement in performance. Additionally, the selection of data augmentation strategies relies on practical scenarios, especially dataset sizes. Among configurations with fitted n0.95 values in the very small sample range (approximately 20–30 original image patches), the mixed all-in configuration frequently showed a favourable trade-off; among configurations with fitted sample requirements of around 50–100 patches, rotation also showed favourable performance.