VGG19-Based Neural Style Transfer for Data Augmentation in Lung Cancer Detection

Deep learning models have shown promising results in computational pathology, but their high data demand poses challenges for multiinstitutional data collaborations due to privacy concerns. Federated learning offers a novel approach to collaboratively train models across different sites without sharing private data. However, imbalanced data distributions among participating sites can lead to performance degradation and instability in federated learning. This study proposes a federated version of the neural style transfer algorithm, originally introduced by Gatys et al., as a data augmentation technique for highly class-imbalanced lung cancer imaging datasets. The proposed method involves selecting characteristic style images and generating gram style matrices at local sites, which are then transferred to other imbalanced sites without leaking any private data. This technique aims to augment the underrepresented classes and mitigate the effects of data imbalance. The proposed method utilizes the VGG19 architecture, a powerful convolutional neural network, as the backbone for feature extraction and style transfer, leveraging its deep layers to capture both content and style information effectively. The approach was evaluated on a federated learning configuration using a lung cancer imaging dataset from multiple institutions.