MICCAI 2018 Daily - Tuesday

tagged, for some type of finding, immediately made all the reports with that sentence as positive for that finding. They were able to cover a lot of ground by covering only 20,000 sentences, rather than going over the million reports they had in the dataset. This was the approach that they used to get the labels for the training. Of course, there were a lot of sentences that their human annotators did not read. What they found out is that those 20,000 sentences were enough to collect the labels. They had a lot of studies where there might have been sentences that they felt were also a positive finding, that they couldn’t detect because they weren’t included in the 20,000 sentences. When they added those studies into the training set, the performance only increased. His take home from this was that it’s better to add more studies to the training set than to have perfect labels: embrace the noise in the labels. What about algorithmic techniques? Jonathan explains: “ Like the rest of the world, we’re using deep learning, we apply a well-known convolutional neural network, DenseNet121, to each of the two images – the frontal view and the lateral view – and then we concatenate their outputs at the end. From there, we continue by making the predictions for the 40 findings. ” In terms of next steps, Jonathan tells us they are looking to extend this study to other modalities and are currently working on doing this with CT images. To find out more, visit Jonathan’s poster [T-27] today at 11:30-12:30 and his oral at 17:45. “ It’s better to add more studies to the training set than to have perfect labels: embrace the noise in the labels! ” 6 Tuesday Oral Presentation