Fig. 5.

Training progression of AgeNet2D using “raw” or “unsegmented” knee MRIs (left) vs. using masked images (right). The mean squared error loss is plotted for the known data (blue), i.e., the training set, and the unknown data (orange), i.e., the validation set, over multiple epochs. Using masked knee MRIs improved training progression and avoided the generalization gap created by CNN training on “raw” images. This supports the assumption that the age estimation problem could be simplified by extracting age-relevant structures from knee MRIs (adapted from [37])