Constrained Generative Adversarial Network Ensembles for Sharable Synthetic Data Generation
The sharing of medical imaging datasets between institutions, and even inside the same institution, is limited by various regulations/legal barriers. Although these limitations are necessities for protecting patient privacy and setting strict boundaries for data ownership, medical research projects that require large datasets suffer considerably as a result. Machine learning has been revolutionized with the emerging deep neural network approaches over recent years, making the data-related limitations even a larger problem as these novel techniques commonly require immense imaging datasets. This paper introduces constrained Generative Adversarial Network ensembles (cGANe) to address this problem by altering the representation of the imaging data, whereas containing the significant information, enabling the reproduction of similar research results elsewhere with the sharable data. Accordingly, a framework representing the generation of a cGANe is described, and the approach is validated for the generation of synthetic 3D brain metastatic region data from T1-weighted contrast-enhanced MRI studies. For 90% brain metastases (BM) detection sensitivity, our previously reported detection algorithm produced on average 9.12 false-positive BM detections per patient after training with the original data, whereas producing 9.53 false-positives after training with the cGANe generated synthetic data. Although the applicability of the introduced approach needs further validation studies with a range of medical imaging data types, the results suggest that the BM-detection algorithm can achieve comparable performance by using cGANe generated synthetic data. Hence, the generalization of the proposed approach for various modalities may occur in the near future.