Generalizable Multi-Site Training and Testing Of Deep Neural Networks Using Image Normalization

The ability of medical image analysis deep learning algorithms to generalize across multiple sites is critical for clinical adoption of these methods. Medical imging data, especially MRI, can have highly variable intensity characteristics across different individuals, scanners, and sites. However, it is not practical to train algorithms with data from all imaging equipment sources at all possible sites. Intensity normalization methods offer a potential solution for working with multi-site data. We evaluate five different image normalization methods on training a deep neural network to segment the prostate gland in MRI. Using 600 MRI prostate gland segmentations from two different sites, our results show that both intra-site and inter-site evaluation is critical for assessing the robustness of trained models and that training with single-site data produces models that fail to fully generalize across testing data from sites not included in the training.