医药魔方 / 最新资讯 / 正文

Nature子刊:脸部拍照可以诊断遗传病?准确率可达91%

偶数 偶数 来源:医药魔方
2019-01-09
Nature
原文

使用手机上的应用程序诊断罕见且难以治疗的遗传病症听起来像纯粹的科幻小说。但是美国FNDA公司的首席科学家Gurovich及其同事,已经将这个概念变为现实。

 

近日,国际顶级学术期刊《自然-医学》发表了一项使用面部图像分析检测遗传性疾病的文章,FDNA公司的研究人员通过使用17000名患者的脸部图像数据集训练了一套计算机深度学习算法,通过这套算法可以帮助诊断遗传性疾病,正确率可达91%,超过了很多专家和临床医生。


1、可通过拍摄面部照片诊断遗传综合征的APP

 

及时诊断遗传综合征可改善预后。但是目前已知的遗传综合征超过8000种,患者可能出现的遗传综合征数量众多且罕见,要做出正确的诊断需要的时间极其漫长,并且花费昂贵。另外,对于非经典表现综合征或超典型综合征的诊断受临床专家先前经验的限制,这使得计算机系统作为参考越来越重要。

 

该研究指出,8%的人群可能患有遗传综合征,并且许多人具有可识别的面部特征。例如,Angelman综合征,一种影响神经系统的疾病,很多典型的面部特征,例如牙齿间距较宽、斜视、舌头突出等。因此,基于面部特征识别遗传综合征便成为可能。

早期的计算机辅助综合征识别技术通过对患者面部图像的分析显示出了帮助临床医生诊断的希望。但是关于这种可能性的研究所采用的训练数据集规模不大,仅能识别少量综合征。

 

位于波士顿的FDNA公司的研发人员Gurovich及其同事开发了一个APP——Face2Gen。这个APP基于一种AI技术DeepGestalt TM(新型面部图像分析框架),该框架使用计算机视觉和深度学习算法,可成功识别数百种遗传性疾病的面部表型。Gurovich和他的团队通过使用来自诊断出200多种不同遗传综合征的患者数据库中的17,000名患者的面部图像来训练DeepGestalt 。


2、准确率可达91%


在新的研究论文中,研究人员深入解释了这种技术的工作原理。首先输入面部图像,使用基于DCNN(深度卷积神经网络)的级联方法进行人脸检测,将输入的图像裁剪成多个面部区域,将每个区域馈入DCNN以获得softmax函数(柔性最大值传输函数),指示其与模型中每个综合症的对应关系。然后合计并将所有区域中DCNN的输出函数进行分类,以获得遗传综合征的最终排序列表。右侧的直方图表示DeepGestalt输出的遗传综合征,按合计的相似性分数排序。

 


该团队发现,在确定502张所选面部图像中遗传综合征的两组不同测试中,这项AI技术明显胜过临床医生。在每次测试中,通过这项AI生成的潜在的综合症列表,其前10项建议中标出了正确的遗传综合征的概率为91%。

 

另一项测试旨在确定Noonan综合征的不同遗传亚型,该综合症具有一系列独特的特征和健康问题,如心脏缺陷。在这项测试中,AI技术深度学习算法的成功率为64%,在先前的研究中,观察Noonan综合征患者图像的临床医生仅能确定20%的病例。但是研究人员也表示,AI技术只能作为一种辅助手段,结果需要医生进行最终确定。


3、谨防滥用风险


该论文的共同作者Karen Gripp表示,这篇论文的重要性在于详细描述了如何训练算法及其工作原理。虽然还有其他同类系统,但没有一种系统有这么多的案例和疾病可以分析。本文为与其他系统进行比较创建了标准,并且为使用该工具用于其他研究提供了参考。

 

Gripp希望下一步能够利用该技术分析面部的侧视图,侧视图也可以成为诊断时的有用信息。她还希望获得更多关于不同种族背景的数据,因为绝大多数上传的面孔都是欧洲人。但是,她指出,该技术在不同种族中表现良好。目前FNDA正在开发使用该技术的嵌入式解决方案,可以授权给其他医疗保健和技术组织,以便其在自己的平台使用该技术。

 

另外研究人员也承认,该技术存在一定的风险。因为拍摄脸部太容易了,因此,该技术有可能被雇主或保险公司滥用。他们表示,对DeepGestalt等工具的分配和使用进行适当的监管至关重要。


原文:Identifying facial phenotypes of genetic disorders using deep learning

作者:Yaron Gurovich , Yair Hanani , Karen W. Gripp,et al

机器翻译

Diagnosing rare and difficult-to-treat genetic conditions using apps on mobile phones sounds like pure science fiction.But Gurovich, chief scientist at FNDA, and colleagues in the US, have turned the concept into reality.

Recently, the world's top academic journal Nature-Medical published an article using facial image analysis to detect hereditary diseases. Researchers at FDNA trained a computer deep learning algorithm by using a facial image dataset of 17,000 patients. This algorithm can help diagnose hereditary diseases with a accuracy rate of 91%, which exceeds that of many experts and clinicians.

1, APP

which can diagnose genetic syndromes by taking facial photographs; timely diagnosis of genetic syndromes can improve prognosis.However, there are more than 8,000 known genetic syndromes. The number of genetic syndromes that may occur in patients is numerous and rare. It takes a long time to make a correct diagnosis and is expensive.In addition, the diagnosis of non-classical manifestations or supertypical syndromes is limited by prior experience of clinical experts, which makes computer systems increasingly important as a reference.

The study states that 8% of the population may have a genetic syndrome and that many have identifiable facial features.For example, Angelman syndrome, a disease affecting the nervous system, has many typical facial features, such as wide spacing of teeth, strabismus, protruding tongue, etc.Therefore, it is possible to identify genetic syndromes based on facial features.

Early computer-aided syndrome recognition techniques have shown promise in helping clinicians diagnose by analyzing patients' facial images.But the training data set used by studies on this possibility is small and only identifies a small number of syndromes.

R & D staff at FDNA, based in Boston, Gurovich and colleagues developed an app — Face2Gen.This app is based on an AI technology, DeepGestaltTM (New Facial Image Analysis Framework), which uses computer vision and deep learning algorithms to successfully identify the facial phenotypes of hundreds of genetic diseases.Gurovich and his team trained DeepGestalt by using facial images from 17,000 patients in a database that diagnosed more than 200 different genetic syndromes.

2, accuracy up to 91%

In the new research paper, the researchers explained in depth how this technique works.First the facial image is input, face detection is performed using a cascading method based on DCNN (depth convolution neural network), the input image is cropped into multiple facial regions, and each region is fed into DCNN to obtain a softmax function (flexible maximum transfer function) indicating its correspondence with each syndrome in the model.The output functions of DCNN in all regions are then summed and classified to obtain a final ranked list of genetic syndromes.Histograms on the right indicate the genetic syndromes output by DeepGestalt, sorted by the summed similarity scores.

The team found that this AI technique significantly outperformed clinicians in two different sets of tests for determining genetic syndromes in 502 selected facial images.In each test, the list of potential syndromes generated by this AI has a 91% probability of identifying the correct genetic syndrome in the top 10 recommendations.

Another test aims to identify different genetic subtypes of Noonan syndrome with a unique set of features and health problems such as heart defects.In this test, the AI technique deep learning algorithm had a success rate of 64%, and in a previous study, clinicians observing images of patients with Noonan syndrome were able to identify only 20% of casesHowever, the researchers also said that AI technology can only be used as an aid, and the results need to be finalized by doctors.

3, beware of abuse risk

Karen Gripp, co-author of the paper, said that the importance of this paper lies in the detailed description of how to train the algorithm and how it works.Although there are other similar systems, no system has so many cases and diseases to analyze.This paper creates standards for comparison with other systems and provides a reference for using this tool for other studies.

Gripp hopes that the next step is to use this technique to analyze the lateral view of the face, which can also be useful information at the time of diagnosis.She also wants more data on ethnic backgrounds, since the vast majority of uploaded faces are European.However, she points out that the technique performs well in different races.Currently FNDA is developing embedded solutions that use the technology and can be licensed to other healthcare and technology organizations to use the technology on their own platforms.

In addition, the researchers also acknowledged that the technology has certain risks.Because it is too easy to shoot the face, the technology could be misused by employers or insurance companies.They say proper regulation of the distribution and use of tools such as DeepGestalt is essential.

Original: Identifying facial phenotypes of genetic disorders using deep learning

Author: Yaron Gurovich, Yair Hanani, Karen W.Gripp, et al

扫码实时看更多精彩文章

版权及免责声明 本文由医药魔方原创,版权归医药魔方所有。未经许可,严禁任何媒体或个人以任何形式摘编、改写、复制、转载本文内容。对于恶意侵权行为,医药魔方保留采用法律手段追究的权利。媒体内容及商务合作请联系医药魔方工作人员(微信号:medicube)

魔方微信公众号