Deploying Deep Learning into Practice: A Case Study on Fundus Segmentation

We study the deployment of a deep learning medical image segmentation pipeline, which sees new input data, not contained in the training and evaluation database. Like in application, although this data shows the same properties, it may stem from a slightly different distribution than the training set because of differences in the hardware setup or environmental conditions. We show that, although cross-validation results suggest high generalization, segmentation score drops significantly with pre-processed data from a new database. The positive effects of a short fine-tuning phase after deployment, which seems to be necessary under such conditions, can be observed. To enable this study, we develop a segmentation pipeline comprising pre-processing steps to homogenize the data contained in 4 databases (DRIONS, DRISHTI-GS, RIM-ONE, REFUGE) and an artificial neural network (NN) segmenting optic disc and cup. This NN can be trained using exactly the same hyperparameters on all 4 databases, while achieving performance close to state-of-the-art methods specifically designed for the individual databases. Furthermore we deduct a hierarchy of the 4 databases with respect to complexity and broadness of contained samples.