Fig. 5:

Models Genesis enable better optimization than learning from scratch, evident by the learning curves for the target tasks of reducing false positives in detecting lung nodules (NCC) and pulmonary embolism (ECC) as well as segmenting lung nodule (NCS), liver (LCS), and brain tumor (BMS). We have plotted the validation performance averaged by ten trials for each application, in which accuracy and dice-coefficient scores are reported for classification and segmentation tasks, respectively. As seen, initializing with our pre-trained Models Genesis demonstrates benefits in the convergence speed.