Radiology
1. Deep Learning for Triage of Chest Radiographs: Should Every Institution Train Its Own System? （paper）
文章貢獻(xiàn)：用三個(gè)well-known網(wǎng)絡(luò)：AlexNet, ResNet-18和DenseNet-121對(duì)200000張胸部X光進(jìn)行分類，用ImageNet預(yù)訓(xùn)練的權(quán)重聊浅。DenseNet-121性能最好(AUC=0.96)褥紫，AlexNet最差，但是不同網(wǎng)絡(luò)之間的差距很小萄窜。經(jīng)典的機(jī)器學(xué)習(xí)算法SVM也做了比較圣勒，效果也還不錯(cuò)(AUC=0.93)膝迎，但是substantially inferior to CNN.
How could one use such a system? 作者建議：It could be used for triage in areas without access to trained radiologists and for work?ow prioritization in clinics with sta? shortages.
算法+人的組合性能最好：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).
訓(xùn)練數(shù)據(jù)對(duì)網(wǎng)絡(luò)性能的影響： 超過(guò)2W后性能沒(méi)顯著提升盆色。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 di?erence 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. 就該任務(wù)而言，單個(gè)機(jī)構(gòu)搜集2W左右的數(shù)據(jù)就能訓(xùn)練一個(gè)高性能的網(wǎng)絡(luò)祟剔，收集這么多數(shù)據(jù)對(duì)大多數(shù)機(jī)構(gòu)都是可行的隔躲。
這是引人深思的觀點(diǎn)（thought-provoking statement）。這違背了（run counter）了本期刊的notion以及 the regulatory authorities物延。通常我們要求算法必須在大的多中心的數(shù)據(jù)集上驗(yàn)證宣旱。
編輯的期望：不僅僅給出二值的預(yù)測(cè)結(jié)果，還要檢測(cè)出圖像中有不正常的區(qū)域叛薯。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.
