MIDL Vision 2020
the network. This gives us scalability of the network. If patient-level training data from a new hospital is available, and we can only update the weights for the patient-level training stage, we can adapt our framework to that hospital’s data to fully customize our network to give great performance on their data.” Thinking ahead, Luyao says there are technical elements of the work that he would like to improve. In stage I, it uses ResNet , which could be replaced by a more efficient network, like DenseNet or another more advanced architecture. Also, in stage I, the balance between the combination of classification and attention loss couldbe furtheroptimized. Oral Presentation 8 “… we can adapt our framework to that hospital’s data to fully customize our network to give great performance on their data!” In Stage II training, which is the patient- level training, the weight of the stage I image encoder networks is fixed, but he thinks if they use end-to-end training in stage II, they can then update the weights of the stage I network. This might improve the results. On behalf of all of our readers, we offer Luyao our sincerest congratulations on completing his PhD and wish him all the best for a successful future! To learn more about Luyao’s work [O042], you are invited to visit Oral Session #6 - Attention at 12:30-13:30 and Poster Session #5 at 13:30-15:00 today. An example of the attention maps that can provide PE visual localization.
Made with FlippingBook
RkJQdWJsaXNoZXIy NTc3NzU=