编程知识 cdmana.com

At the top of the meeting, two papers were published successively, and Huawei cloud medical AI emerged in a low profile

Abstract :2020 International Conference on medical image computing and computer aided intervention (MICCAI 2020), The results of the paper reception have been published . Huawei cloud Medical AI The team cooperated with Huazhong University of science and technology 2 Research results were selected .

At the same time, two research results were included in the industry summit , Huawei cloud Medical AI Layout , Low key comes to the surface .

2020 International Conference on medical image computing and computer aided intervention (MICCAI 2020), The results of the paper reception have been published . among , Huawei cloud Medical AI The team cooperated with Huazhong University of science and technology 2 Research results were selected .

MICCAI 2020 It spans two fields of medical image computing and computer-aided intervention , There has been a 16 The history of its development is , Is internationally recognized as the industry's top academic conference . It has not only international influence and academic authority , It is also a hot wind vane in the field of medical image analysis , It is also the place to verify the gold content of relevant research results .

semantics / Case segmentation is a hot research topic in the field of medical image computing in recent years ,70% The above international competitions are all around it . This time Huawei cloud Medical AI Team 2 Papers , It aims at the semantics in the field of medical image computing / Instance segmentation problem , Explore how the heart 、 The liver 、 The original image data of prostate and other organs are transformed into high-value spatial structured information . It has a strong clinical value , Not only can it help doctors make decisions , It can also help doctors complete preoperative planning , Tumor dynamic monitoring and other tasks .

In the research paper , Huawei cloud Medical AI The approach proposed by the team , Used to solve imaging by medical devices 、 The edge of the object to be segmented is not clear due to the structure of the organ focus itself and other factors —— This is a scene where deep learning algorithms are hard to work before .

Among the two methods proposed by Huawei cloud team , Each one shows an effect beyond the traditional method .

Medical care +AI Latest achievements

One of the papers , be known as “Pay More Attention to Discontinuity for Medical Image Segmentation”, The research is how to segment the discontinuous area in medical image ( Such as organ lesions ).

The paper says , The existing segmentation methods deal with this situation , The discontinuous position in the region is often misjudged as the boundary of the region , This leads to inaccurate regional boundaries .

For example, in the picture below ( On the left is the label map , On the right is the existing method of image segmentation , The yellow circle is the missing part ):

In this paper , Huawei cloud Medical AI Huazhong University of science and technology , The concept of inaccurate edge segmentation caused by discontinuity in region is discussed , And put forward the solution : Raise the attention of the discontinuous position .

say concretely , It's the application of edge detectors to identify discontinuities , And “ Discontinuous ” Supervisory signals are added to loss In the objective function , Routine coordination Dice loss Combined into a multi task objective learning function , In this way, we can make more accurate edge recognition , The algorithm framework is shown in the figure below :

They tested the algorithm in three medical image segmentation tasks , Namely :MRI Heart segmentation dataset -Cardiac500、MRI Prostate segmentation dataset T2-SPIR and MRI Liver segmentation dataset Medical Segmentation Decathlon.

Results show , Compared with the existing baseline method , The core indicators of segmentation results have been improved . among . On the heart segmentation and migration task Cardiac500 Migrate to ACDC The results have improved 5.1 percentage .

In order to further verify the effectiveness of their proposed method , They further analyzed Cardiac500 Data set 2645 Distribution of segmentation results of test samples , The results show that the core indicators less than 0.8 The sample of , by comparison , Baseline methods are 13 Samples are lower than 0.8.

Another paper , titled “Learning Directional Feature Maps for CardiacMRI Segmentation”, The same is Huawei cloud healthcare AI The team combined with the research results of Huazhong University of science and Technology .

Usually , Factors such as inhomogeneous magnetic field and visceral motion during MRI can produce artifacts , Blur the boundary of the target . However, the current segmentation methods based on deep learning are lack of effective semantic pixel level association , As a result, the segmented object cannot maintain the anatomical structure , As shown in the figure below :

This is Huawei cloud healthcare AI What the team is going to solve in this paper , They put forward a feature map by learning direction , Strengthen the semantic level association between pixels , By increasing class spacing , Narrowing the class spacing , To maintain the anatomical structure of an object , To achieve high-precision edge segmentation . The specific process is shown in the figure below :

First , use U-Net To learn the initial segmentation rendering . after , be based on U-Net The trunk , adopt DF The module learns the intensity information and direction information of each pixel direction field .

Next , The direction field information obtained by learning is used to modify the initial segmentation effect iteratively , Using the results of organ middle segmentation to guide edge segmentation . Last , Joint initial segmentation effect + Direction field learning + Revise the segmentation effect and other tasks for multi task learning .

This paper shows the segmentation and generalization performance of this method . Compared with the existing methods , It's on the heart segmentation and migration task (Self-collected ->ACDC, ACDC ->Self-collected) They have been promoted separately 1.1 Points and 1.7 A little bit .

Based on the hot research topics of the industry , Both papers were included in the industry summit , Huawei cloud Medical AI You can see the strength of .

Based on the above two methods , Huawei cloud Medical AI The team and Huazhong University of science and technology jointly developed a set of heart based on deep learning AI service , It can realize the automatic segmentation of heart structure 、 And accurate quantitative analysis , Realize the second level output of single case quantitative results ,AI+ The overall efficiency of doctor review is tens of times faster than that of pure manual quantitative evaluation . at present , The service has been successfully launched in Huawei cloud .

But it's only part of the research in recent years , Huawei cloud is in medical treatment AI field , It's been a long time , Especially in the field of medical imaging .

Huawei cloud Medical AI Layout emerges

Judging from the research results , in fact , In last year's MICCAI as well as MICCAI-MIML On , Huawei cloud Medical AI The team already has 3 Papers are shortlisted , Cervical cancer screening coverage 、 Application scenarios such as stroke segmentation and automatic generation of plain film diagnosis report .

In recent years, many medical imaging related AI At the challenge , Huawei's cloud technology strength has also reached the world's leading level .

For example Grand-Challenge Fetal ultrasound head circumference measurement competition (HC18) On , Huawei cloud surpasses the Chinese University of Hong Kong 、 Chinese academy of sciences 、 Queen's University of Canada, etc 100 Multiple universities and research institutions , With 1.89mm The average absolute error of is the first .

Some time ago , We reported on IEEE Fellow、AI Daniel Tianqi joined Huawei as the chief scientist in the field of cloud artificial intelligence .

Tian Qi is a big bull in the field of computer vision , The dominant AI Cutting edge research in vision , After he joined , It will surely enhance Huawei cloud's basic research strength in the field of computer vision . Predictably, , After Tian Qi joined , Huawei cloud Medical AI, Especially in medical imaging , There will be more progress in the future .

But it's not just research , Huawei cloud is still actively exploring how to AI Technology is coming to the ground fast .

The past few years , They work with companies in the healthcare industry and hospitals and Universities , Provide end-to-end AI Enabling platform , Push AI Applied to industry scenarios .

2019 year 6 month , Huawei cloud and Jinyu medical cooperation , stay AI Breakthroughs have been made in the application and development of auxiliary pathological diagnosis . They trained a cervical cancer screening model , The rate of negative discharge is higher than 60% On the basis of , The correct rate of negative film interpretation is higher than 99%, meanwhile , The detection rate of positive lesions is more than 99.9%.

At one stroke, it has been published internationally AI The highest level of adjuvant cervical cancer screening . And in the diagnosis speed also greatly improved : Each case of pathological interpretation only needs 36 second , It's manual interpretation 10 times .

During the epidemic , Huawei cloud and blue net technology and other partners , Created a medical imaging aided diagnosis system based on artificial intelligence , Will diagnostic efficiency from the past 10-15 Minutes become 10-15 second , Greatly alleviated the medical pressure .

With years of technology accumulation , Huawei cloud has launched enterprise level medical imaging for the industry AI platform , Support the end-to-end traceability of the whole process AI modeling , Help medical imaging AI More systematic 、 Fast 、 Go to the market safely .

Besides , In genomics and pharmaceuticals , Huawei cloud also has a lot of layout and accumulation .

Soon after the outbreak of the new epidemic this year , Huawei cloud and its partners have formed a joint research team , Based on Huawei cloud medical agent platform (EIHealth), For all the new coronavirus 21 Target proteins for computer-assisted drug screening . Tens of millions of simulations have been completed in just a few hours , And published the research results in time , It provides support for the global anti-virus research and development .

And before , Such a large-scale calculation often takes months to complete .

While the global anti epidemic situation is still grim , This help shows that AI The inclusive side , This is exactly what Huawei cloud does to its medical treatment AI Look forward to of : Solve the fundamental problems in the medical field , adopt AI Technology conversion , Prevention of human diseases 、 The diagnosis 、 Healing contributes .

But in computing power 、 The algorithm and application platform are mature , Huawei cloud Medical AI The speed of advance is also accelerating .

Huawei cloud medical agent platform (EIHealth) It has been opened to the outside world , If you're interested , You can visit the page for details → Portal .

This article is shared from Huawei cloud community 《 Hua Weiyun Medical Co., Ltd AI: Two papers will be published in succession , Research and landing acceleration 》, Original author : Meat worms .

 

Click to follow , The first time to learn about Huawei's new cloud technology ~

版权声明
本文为[Huawei cloud developer community]所创,转载请带上原文链接,感谢
https://cdmana.com/2020/12/20201224120550855S.html

Scroll to Top