Radiology
1. Deep Learning for Triage of Chest Radiographs: Should Every Institution Train Its Own System? （paper）
文章贡献：用三个well-known网络：AlexNet, ResNet-18和DenseNet-121对200000张胸部X光进行分类，用ImageNet预训练的权重。DenseNet-121性能最好(AUC=0.96)，AlexNet最差，但是不同网络之间的差距很小。经典的机器学习算法SVM也做了比较，效果也还不错(AUC=0.93)，但是substantially inferior to CNN.
How could one use such a system? 作者建议：It could be used for triage in areas without access to trained radiologists and for workflow prioritization in clinics with staff shortages.
算法+人的组合性能最好：Alternatively, the output of the network can be averaged with a rating provided by a human reader. Dunnmon et al showed that such a combined human and artificial intelligence system achieves an AUC of 0.98, which is significantly better than the computer system alone, and achieves a higher accuracy than human reading alone (the best network alone still has slightly lower accuracy than human reading).
训练数据对网络性能的影响： 超过2W后性能没显著提升。Experimental results with training sets of 2000, 20 000, and 200 000 images are compared. Results when using only 2000 images are substantially worse, but the difference between 20 000 and 200 000 training images is insignifcant, as measured in a hold-out test set of 1000 images that were carefully reannotated by expert readers. 就该任务而言，单个机构搜集2W左右的数据就能训练一个高性能的网络，收集这么多数据对大多数机构都是可行的。
这是引人深思的观点（thought-provoking statement）。这违背了（run counter）了本期刊的notion以及 the regulatory authorities。通常我们要求算法必须在大的多中心的数据集上验证。
编辑的期望：不仅仅给出二值的预测结果，还要检测出图像中有不正常的区域。It would be advisable to train systems not only to provide a binary output label but also to detect specifc regions in the images with specifc abnormalities. This would require annotation of such regions on many training images. Tis would be a step toward a deep-learning network that explains to the user why it arrived at the overall conclusion that examination results might be abnormal.
