Generative Bone Lesions Synthesis for Data Augmentation in X-ray

Insufficient training data and severe class imbalance are often the main limiting factors when developing machine learning models for the classification of rare diseases. In this work, we address this problem by augmenting the training set with synthesized images. We pose the generative task as an unsupervised imagepatch translation problem with the aim to generate bone lesions on images without pathology. In experimental results, we show that this can enable the training of superior classifiers achieving better performance on a held-out test set in the binary classification task of bone lesion detection. Additionally, we demonstrate the feasibility of transfer learning and apply a generative model that was trained on one bone to another.