MIDL Vision 2020

Luyao Shi 7 Luyao tells us the novelty of this work isn’t only about the scale of the data, but its technical innovations too: “Manual annotation of data is very laborious. That’s a general problem for medical imaging applications using deep learning. People will often use patient-level labelled data to provide training data. This means the network can be trained in an end-to- end fashion, but once we analyze the attention maps, we can find that they only cover the most discriminative regions and not necessarily the lesions. This causes a problem because the lesions sometimes have correlations with background regions like the ribs, and in that case it’s easy for the network to focus on those regions instead. Even though it can still show a high accuracy, if you look at the attention map, you can see that it is completely focused on the wrong regions. Then when it comes to patient-level, because of the many false positives , you will have poor resul ts.” This work uses supervised attention training to help the network focus its attention on pulmonary embolism. Later, the network can be used as an image encoder to greatly improve the final patient-level prediction results. One benefit of their proposed framework is its ability to provide localized attention maps that indicate possible PE lesions, which could potentially help radiologists accelerate the diagnostic process. Luyao explains the scalability of the network is also important: “In stage I we use attention loss to train the image encoder and in stage II we use the patient-level labelled data to train "… potentially help radiologists accelerate the diagnostic process" Proposed hybrid training framework with two stages.