AI Tech Outperforms Human

Author: Zongwei Zhou | 周纵苇
Weibo: @MrGiovanni
Email: zongweiz@asu.edu
Talk in BMI 570 Symposium journal club


Good afternoon. My name is Zongwei Zhou. I'm a 3rd year Ph.D. student in Dr. Jianming Liang's lab. My research interest is the artificial intelligence system in computer vision, especially in medical image analysis. So, no surprise, I'd like to introduce you an impressive paper in our field, yet it has been published for more than two years.

Dermatologist-level classification of skin cancer with deep neural networks
Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017 Feb;542(7639):115.

AI system has recently been shown to exceed human performance in many visual tasks, such as playing Atari games (Human-level control through deep reinforcement learning), Go (Mastering the game of Go with deep neural networks and tree search), and object recognition (Imagenet large scale visual recognition challenge), thanks to advances in computation power and enormous amount of labeled dataset. For example, in the image recognition task, one is asked to identify the object from an image. The red line in the figure denotes the average human performance; AI system did a better job than average human performance in 2016, measured by prediction error rates of unseen images.

AI system is challenging doctors in medical image analysis as well. When we invent an AI system to detect diseases from such medical images as CT, MRI, ultrasound, microscopy, etc., usually we try to compare its performance with that of medical experts. Most of the time, AI systems tend to make a bunch of silly mistakes, but in some specific tasks, it may approximate human experts accuracy and even more. IEEE Spectrum, for example, proposed a survey on the performance between AI system and doctors since May 2015. It depicts a big picture of the recent progress of AI research in medical image analysis. Today, the paper that I will present covers one of the diseases reported in this survey -- skin cancer.

This paper states that the deep convolutional neural networks (CNN) achieve superior performance with all tested experts in both common skin cancer and deadliest skin cancer identification tasks, demonstrating an artificial intelligence capable of classifying skin cancer with a level of competence comparable to human experts. Such a statement, AI tech outperforms human, is usually occurred as the compelling headline in nowadays technical news because it's always an eye-catching topic; yet how to convincingly verify the study design and comparison result remains a challenging problem. In doing so, this study is perhaps among the first to provide a substantial exemplary to convey comprehensive controlled experiments between the deep network and human experts in different perspectives.

A persistent question: toward which types of disease identification, computer-aid diagnosis (CAD) may contribute the most? (1) Time, those disease detection is not difficult but tedious and time-consuming, that not enough number of doctors can cover all the clinical needs on time. (2) Accuracy, those disease detection may be very challenging even for human experts. AI system is trained to identify patterns that are not obvious to humans. But in summary, those diseases should be at least the common disease for now.

There are 5.4 million new cases of skin cancer in the United States every year. One in five Americans will be diagnosed with a cutaneous malignancy in their lifetime. Early detection is critical, as the estimated 5-year survival rate for melanoma drops from over 99% if detected in its earliest stages to about 14% if detected in its latest stages. Thereby, skin cancer is one of the most suitable applications where AI system may contribute its power.

Controlled experiments study design is used in this paper, to compare AI system with human experts. Twenty-one board-certified dermatologists are selected to represent the average performance of human experts. In a technical perspective, nine-fold cross-validation was performed to give robust and statistically meaningful results, and both the AI system and human experts are presented in two tasks: common skin cancer and deadliest skin cancer identification.

They have collected a large clinical dataset for results verification, consisting of a hundred thousand clinical images with two thousand different diseases. AI system is data hungry, requiring a huge labeled dataset. The existing body of AI system uses small datasets of typically less than a thousand images of skin lesions, which, as a result, do not generalize well to new images. So, this study is potential to lead to a statistically meaningful conclusion and has more chance to get FDA approval.

The algorithm behind the AI system in this paper is a deep convolutional neural network, which is the key breakthrough in the field of computer vision for now. One question I found in the discussion section is why use the neural network, rather than other conventional machine learning algorithms? The answer is deep neural network is more powerful when a large amount of labeled data is in hand. Here, the x-axis is the number of data, the y-axis is the potential performance that these algorithms can achieve. Decades ago, as we don't have so many gold standards labeled associate with images and so powerful computation resource, traditional machine learning such as SVM, random forests, Adaboost were very useful, but their performance is upper bounded. The deep network can automatically learn the representation directly from the input image. Here deeper means the length of the sequence of layers.

The deep network appears like a tiramisu; input image goes through a sequence of layers, those layers are regarded as image feature generator. When extracting higher-level representation from the original image, it becomes easy to predict the final class of the input image.

Here display the overall performance between the AI system and 21 human experts. X and Y axes denote sensitivity and specificity. The larger, the better. Each red point represents one human expert performance in terms of sensitivity and specificity, and the green point represents their average performance and standard deviation. The prediction of the network is the probability of the input image belonging to one specific disease.

The deep network intends to model the human brain, that is designed to recognize patterns. The huge successes of deep networks are usually applied in a black box manner because of their nested non-linear structure---no visual information is provided about what exactly makes them arrive at their predictions. To develop more transparent behavior in diseases identification, they visualize and interpret the class activation maps, which reflect which regions contribute the most when the deep network makes its predictions.

Strength
  • Develop a very powerful algorithm that learns from the image directly
  • Visualize the salience map that may be helpful for medical education
Limitation
  • Annotating a taxonomy of skin sub-class demands costly, specialty-oriented knowledge and skills.

For skin cancer, thanks to the research group in Stanford, a large number of the labeled database has been organized. Unfortunately, one should redo this process again and again when applying this technic to other diseases.

This paper demonstrates the effectiveness of deep learning in dermatology. A deep convolutional neural network, trained on general skin lesion classification dataset, matches the performance of at least 21 board-certified dermatologists. This fast, scalable method is deployable on mobile devices and holds the potential for substantial clinical impact.

I thank you!

References

[1] Do machines actually beat doctors?

最后编辑于
©著作权归作者所有,转载或内容合作请联系作者
  • 序言:七十年代末,一起剥皮案震惊了整个滨河市,随后出现的几起案子,更是在滨河造成了极大的恐慌,老刑警刘岩,带你破解...
    沈念sama阅读 116,857评论 1 238
  • 序言:滨河连续发生了三起死亡事件,死亡现场离奇诡异,居然都是意外死亡,警方通过查阅死者的电脑和手机,发现死者居然都...
    沈念sama阅读 50,582评论 1 199
  • 文/潘晓璐 我一进店门,熙熙楼的掌柜王于贵愁眉苦脸地迎上来,“玉大人,你说我怎么就摊上这事。” “怎么了?”我有些...
    开封第一讲书人阅读 71,827评论 0 163
  • 文/不坏的土叔 我叫张陵,是天一观的道长。 经常有香客问我,道长,这世上最难降的妖魔是什么? 我笑而不...
    开封第一讲书人阅读 35,291评论 0 126
  • 正文 为了忘掉前任,我火速办了婚礼,结果婚礼上,老公的妹妹穿的比我还像新娘。我一直安慰自己,他们只是感情好,可当我...
    茶点故事阅读 42,068评论 1 204
  • 文/花漫 我一把揭开白布。 她就那样静静地躺着,像睡着了一般。 火红的嫁衣衬着肌肤如雪。 梳的纹丝不乱的头发上,一...
    开封第一讲书人阅读 34,883评论 1 124
  • 那天,我揣着相机与录音,去河边找鬼。 笑死,一个胖子当着我的面吹牛,可吹牛的内容都是我干的。 我是一名探鬼主播,决...
    沈念sama阅读 26,856评论 2 205
  • 文/苍兰香墨 我猛地睁开眼,长吁一口气:“原来是场噩梦啊……” “哼!你这毒妇竟也来了?” 一声冷哼从身侧响起,我...
    开封第一讲书人阅读 26,012评论 0 119
  • 想象着我的养父在大火中拼命挣扎,窒息,最后皮肤化为焦炭。我心中就已经是抑制不住地欢快,这就叫做以其人之道,还治其人...
    爱写小说的胖达阅读 24,971评论 5 169
  • 序言:老挝万荣一对情侣失踪,失踪者是张志新(化名)和其女友刘颖,没想到半个月后,有当地人在树林里发现了一具尸体,经...
    沈念sama阅读 29,049评论 0 177
  • 正文 独居荒郊野岭守林人离奇死亡,尸身上长有42处带血的脓包…… 初始之章·张勋 以下内容为张勋视角 年9月15日...
    茶点故事阅读 26,280评论 1 167
  • 正文 我和宋清朗相恋三年,在试婚纱的时候发现自己被绿了。 大学时的朋友给我发了我未婚夫和他白月光在一起吃饭的照片。...
    茶点故事阅读 27,513评论 1 173
  • 白月光回国,霸总把我这个替身辞退。还一脸阴沉的警告我。[不要出现在思思面前, 不然我有一百种方法让你生不如死。]我...
    爱写小说的胖达阅读 21,876评论 0 24
  • 序言:一个原本活蹦乱跳的男人离奇死亡,死状恐怖,灵堂内的尸体忽然破棺而出,到底是诈尸还是另有隐情,我是刑警宁泽,带...
    沈念sama阅读 24,476评论 2 163
  • 正文 年R本政府宣布,位于F岛的核电站,受9级特大地震影响,放射性物质发生泄漏。R本人自食恶果不足惜,却给世界环境...
    茶点故事阅读 28,266评论 3 171
  • 文/蒙蒙 一、第九天 我趴在偏房一处隐蔽的房顶上张望。 院中可真热闹,春花似锦、人声如沸。这庄子的主人今日做“春日...
    开封第一讲书人阅读 23,245评论 0 4
  • 文/苍兰香墨 我抬头看了看天上的太阳。三九已至,却和暖如春,着一层夹袄步出监牢的瞬间,已是汗流浃背。 一阵脚步声响...
    开封第一讲书人阅读 23,385评论 0 112
  • 我被黑心中介骗来泰国打工, 没想到刚下飞机就差点儿被人妖公主榨干…… 1. 我叫王不留,地道东北人。 一个月前我还...
    沈念sama阅读 29,469评论 2 185
  • 正文 我出身青楼,却偏偏与公主长得像,于是被迫代替她去往敌国和亲。 传闻我的和亲对象是个残疾皇子,可洞房花烛夜当晚...
    茶点故事阅读 29,897评论 2 187

推荐阅读更多精彩内容