Problems with Deep Learning Application to Medical Data: Automatic Segmentation of Corneal Endothelium Layer

Machine learning is treated nowadays as a Holy Grall that solves all problems. However, it is difficult to deny a large area of application and high capabilities to deal with difficult problems, one should have in mind that there are still many troubles which should be considered consciously while preparing experiments. In this work, we investigate the generalizability of the U-Net network for automatic endothelial cell delineation. There are two sets of experiments performed, one showing the capabilities when only data from one distribution (one dataset) is considered, and then, for comparison, the model is trained on one dataset while evaluated with other data sets. The scores measured with the F1 score are very high and comparable in both areas, reaching 95%. However, the visual inspection showed that the numbers should be considered carefully as they do not reflect well the actual state of the generated masks.