Google AI breast cancer detection super human, but flat chest may not be applicable

category:Internet
 Google AI breast cancer detection super human, but flat chest may not be applicable


This scheme is especially suitable for breast cancer that depends on image diagnosis, the incidence is increasing year by year, and the survival rate can be improved by early treatment.

But when the same problem comes to China, the solution is different: in addition to relying on X-ray screening, ultrasound screening is also very important.

Breast cancer is the most common cancer in the world, and it is the leading cause of cancer death in women. In European and American countries, breast cancer accounts for 25% - 30% of female malignant tumors. In China, breast cancer is also the highest incidence of cancer in women.

At the same time, early screening is the most effective way to improve the early diagnosis rate, survival rate and quality of life. Mammography molybdenum target screening, as the most important screening method, has been proved to be effective in reducing breast cancer mortality.

Therefore, to improve the ability of image screening and diagnosis is the focus of breast cancer treatment.

For a long time, the early screening of breast cancer depends on doctors reading films. Even the most senior doctors often have misdiagnosis. False positives lead to unnecessary medical procedures, while false negatives delay treatment.

u25b3 growth and metastasis of breast cancer cells

In addition, with the increasing incidence year by year, medical resources are under pressure. Dominic king, Google Healths UK head, said the UKs Royal College of radiologists estimated in 2018 that the country would need more than 1000 additional full-time radiologists to meet the demand.

AI can play its special role.

Mozziyar etemadi, an assistant professor of Anesthesiology and biomedical engineering at Northwestern University, one of the authors of the paper, said:

Reading mammograms is perfect for machine learning and AI. AI is good at doing the same task over and over again, and then finding something different in 10000 times.

In terms of computing power, the advantages of AI are also highlighted.

Mammograms are so high resolution and data intensive that human eyes (even experienced radiologists) cannot fully process them.

Most hospital computer systems are not functional enough to load all the information provided in mammography, so radiologists can only see the selected information.

Googles algorithm can handle almost all available pixels.

AI performance: the diagnostic rate is significantly improved, and the data sets of Britain and the United States can be used in general

Compared with previous studies, this study has three characteristics: using a large clinical dataset to verify that the same model can be used in the United Kingdom and Britain, and the diagnostic accuracy of AI is significantly higher than that of human beings.

The data set consists of mammograms of more than 76000 women from the UK and more than 15000 women from the US.

The scale of the test set was 25856 British women and 3097 American women.

The researchers tested AI in two different ways:

1. Data sets from the United Kingdom and the United States were used to test and compare the clinical diagnosis rates

In the United States, false positives decreased by 5.7% and false negatives by 9.4%.

Why is the U.S. data so much better than the U.K? It may be related to different clinical diagnosis mechanisms. In the United Kingdom, an X-ray film is first diagnosed by two doctors, and then it is added to the third one when there is an objection; in the United States, it is only diagnosed by one doctor

2. In order to see whether the same model can be used in different populations, the researchers trained the model only with British womens data, and then evaluated it with American womens data set. The results are still considerable:

The false positive rate decreased by 3.5% and the false negative rate decreased by 8.1%.

The study also asked six American radiologists to read 500 X-rays and compare them with AI.

The results are interesting. Although the diagnostic accuracy of AI is higher than that of human, the missed diagnosis of AI and human eye complement each other. AI can catch humans omission, and human can also catch AIs omission.

A the small irregular calcification structure in the figure was recognized by AI, but none of the six radiologists recognized it

The high-density massive malignant tumor in the figure B was recognized by six radiologists, but not by AI

In general, AI captures cancers that are more invasive than doctors. Researchers have yet to explain this.

Noninvasive means that cancer cells will be confined to the breast lobes or breast ducts, while invasive cancer will spread to other parts.

The ultimate goal of this AI system is to assist diagnosis. However, further clinical research is needed before that.

Etemadi, the author of the paper, said:

We just need to better understand when and when tools like AI will help, and ultimately combine technology and human contributions to improve care and its efficiency.

The NYU team has an earlier large-scale study and is open source.

Later, hassbis, the founder of deepmind, the Google research team, replied to Lecun, and we quoted this paper.

When Googles paper introduces relevant research overview, it does involve NYU research twice:

A few studies have described systems that predict breast cancer as having independent performance comparable to that of human experts.

Crucially, the subsequent universal use interval does not exceed 12 months, meaning that more subtle cancer conditions may be overlooked until the next test.

Lecun replied immediately: I was not angry. It was the first time I read the paper that I missed the quotation.

However, he then turned to comments from NYU authors comparing the two studies:

Dont forget that NYU published better results last year, based on more cases, compared with more people reading movies, and the model and data are open-source.

The two studies also sparked more discussion on twitter.

Your main concern is that neither the data nor the model of Google research is public.

Some people say that since it cant be reproduced, it cant be regarded as scientific discovery.

Compare the two studies briefly.

First of all, NYU uses 141473 images of women in the dataset, and Google uses 91000 images of women;

In terms of model performance, the AUC (area under ROC curve) of NYU is 0.895, and that of Google is 0.889 in the UK and 0.895 in the US.

Specificity of early screening of breast cancer in China: ultrasound AI is equally important

For Chinese women, AI screening for breast cancer is a little different, and we need to rely on the Chinese teams own technology.

Because the breast structure of Chinese women is significantly different from that of European and American women. Chinese women have less breast fat, more glands, and 50% of them are fibrous. Glandular obscuration and structural noise were more obvious, and the differentiation of normal breast tissue and lesions was smaller.

Therefore, in addition to X-ray molybdenum target screening, ultrasound screening is also very important.

At present, in the X-ray molybdenum target screening, the early layout of the map based medical and Tencent search, have launched their own AI systems.

Tencents AI system for breast cancer screening is the first relevant system in China. According to the official website data, the sensitivity of calcification and mass detection in the system was 99% and 90.2%, respectively. The sensitivity and specificity were 87% and 96%, respectively.

In addition, Tencent is developing products that combine AI technology with ultrasound screening. According to Zheng Yefeng, a medical AI scientist at Youtu laboratory, the price of ultrasonic instruments is lower, and it has greater potential for promotion at the grassroots level.

Paper portal

https://medium.com/@jasonphang/deep-neural-networks-improve-radiologists-performance-in-breast-cancer-screening-565eb2bd3c9f

https://ieeexplore.ieee.org/document/8861376

reference material:

https://medium.com/@jasonphang/deep-neural-networks-improve-radiologists-performance-in-breast-cancer-screening-565eb2bd3c9f

https://nejmqianyan.cn/article/yxqycp1101540?sg=AbW1NGsHw3NxPd6F

http://html.rhhz.net/ZGZLLC/html/2019-9-1.htm

Oqog7esatvi9gavlwzhwk8hitc6pv0o ufe63 soywd7o5wth1ifqzrpfakdtf6d99rkrxzbojksb8cjyac -- end -- source of this paper: qubit Editor in charge: Liao ziyao, nbjs10040

- Finish -