Computer Vision News - December 2018

or a medical student, we can "train" a human. This intuition is better obtained by having that person do manually the same procedure which the automated system does - looking at the data and the labels. After doing this for a few hours, the human observer will be well trained to look in the right place for the right thing. Once data is collected it must be carefully validated, in order to be sure that the software will work on a valid input. All the research and development work would be made much more difficult (and in many cases impossible) by the use of wrong data. In any case, software results correctness must be validated too. Software should pass through many controls and this is particularly true when input is limited in size. Another key point in this phase is to care for the diversity of data, according to what the software is going to look for. Is it developed to analyse scans coming from different devices or from a single scanner? Will it be used only on adult scans or also on children? Bones of patients looks differently depending on age and even on ethnicity. Training data should be as diverse as the real world data which the algorithm will be fed with in due time. Finally, the test set should somehow differ from the training set, i.e. including different patients. Otherwise the test would not be a real one. Management Project Management Tip 11 Computer Vision News Training data should be as diverse as the real world data which the algorithm will be fed with in due time. Retina of patient with Diabetic Retinopathy. Fundus photo reveals scattered hemorrhages with cottonwool spots and some edema in themacula. If we understand what experts look for in the data, we can feed the algorithm this data as well.