When it comes to the application of artificial intelligence in medical imaging, many people in the industry believe that there are still many constraints and obstacles in the medical field. If AI wants to be implemented in clinical practice, it will be subject to data diversifica

When it comes to the application of artificial intelligence in medical imaging, many people in the industry believe that there are still many constraints and obstacles in the medical field. If AI wants to be implemented in clinical practice, it will be subject to data diversification, complicated labeling, policy clarification and evidence inquiry. It is difficult to understand the limitations of multiple issues such as difficulty and it is difficult to see a clear future at this time.

As a result, many companies choose to retreat or wait and see. Even if they wade into the water, they are only scratching the surface.

"When it comes to medical issues, there is no distinction between China and the United States, technology companies or medical companies. This is a topic related to the life and health of all mankind. When facing this problem, no matter how difficult the process is, someone must stand Come out." said Zhou Xiang, senior technical director of cloud computing at Intel China Data Center.

is in cooperation with many hospitals such as the First Affiliated Hospital of Zhejiang University, Aier Eye Hospital , as well as Alibaba Cloud , School of Mathematics and Physics of Zhejiang University, Zhejiang Deshang Yunxing Image Technology Co., Ltd., Jinhong Technology and Jishi Internet, etc. After several months of joint exploration by various units and enterprises, Intel has confirmed the feasibility of AI in medical imaging-assisted diagnosis of fundus lesions such as thyroid nodules, diabetic retinopathy, and macular degeneration. Therefore, it was decided to increase efforts and fully promote the use of AI technology in Practice in the medical field.

When the medical teams of other companies in the industry focus all their energy on developing their own products, Intel believes that "it is better to teach a man to fish than to teach him to fish." After all, there is a limit to what a company can do. If it uses its own With the technological advantages, platform advantages and industry service capabilities accumulated in the past, supporting more AI medical innovators will be more meaningful than one person working alone.

So how to "teach him to fish"?

After some consideration, Intel decided to provide a more powerful hardware platform and easier-to-use software to teams from all over the world who want to use AI to solve medical problems, allowing them to focus more on solving the problem itself without considering software and hardware limitations. So as to maximize potential energy.

In the end, Intel chose to use the competition as the entry point, and launched a Tianchi Medical Artificial Intelligence Competition with its partners, with the participation of 2887 teams. After 7 months, it officially came to an end recently.

This competition challenges the early screening of pulmonary nodules which is a difficult problem recognized by the medical community.

1. Why choose to start with pulmonary nodule screening?

This competition is initiated and supported by Intel , Alibaba Cloud, and Zero Krypton Technology. It is the first medical artificial intelligence competition in China. Its scale and data volume are also unique in the world. There are 16 well-known domestic cancer hospitals. This competition provided desensitized and annotated low-dose lung CT imaging data of nearly 3,000 high-risk patients.

The reason why pulmonary nodule screening was chosen as the “exam question” is because the situation of lung cancer in my country is becoming increasingly severe. The "2012 China Cancer Registration Annual Report" released by the National Cancer Registration Center as early as 2013 showed that: as of that time, the death rate from lung cancer had increased by 465% in the past 30 years, replacing liver cancer as the leading cause of death from malignant tumors in China.

Some people in the domestic liver cancer identification industry have also pointed out that by comparing the incidence and mortality of lung cancer in China, the United States, and the United Kingdom, the data shows that the incidence of lung cancer in China is lower than that of the United Kingdom and the United States, but the mortality rate has exceeded that of the United Kingdom and the United Kingdom. beautiful. The important reason for this is that when patients are diagnosed with lung cancer, 70% have reached the mid-to-late stage and have missed the best time for diagnosis and treatment.

Therefore, early screening for lung cancer is an urgent and urgent task in China.

Early screening is an important means to reduce lung cancer mortality, and cancer often manifests as pulmonary nodules in the early stages. These pulmonary nodules are very small in size, have low contrast, and have high suppression.

In the past, screening work was completed by radiologists. Since the number of lung CT scans for each patient exceeds 200, the average diagnosis time is more than 20 minutes. When doctors encounter a large number of patients, it is extremely time-consuming and labor-intensive. , and easily missed diagnosis.

At the same time, the imaging manifestations of tiny pulmonary nodules can easily be confused with other tissues or parts, such as capillaries, tuberculosis, pseudotumors, etc., thus disturbing doctors' judgment.

Based on this pain point, imaging departments are in urgent need of new technologies to assist doctors in improving diagnostic efficiency. At this time, artificial intelligence stands out from many technologies. It can learn and imitate doctors' diagnostic "experience" from massive medical imaging data, quickly improve diagnostic capabilities in a short period of time, and assist doctors in reducing misdiagnosis.

2. Software and hardware limitations and solutions faced by the industry before

Although artificial intelligence is updated and iterated rapidly, medical image analysis is a huge engineering problem, and a lot of work must be carried out step by step, especially data forms and characteristics such as medical care are relatively complex. In a complex industry, the problems it faces in artificial intelligence-assisted diagnosis are one after another and closely related to each other:

• First of all, GPU is difficult to bear the heavy burden, but Xeon Phi has unique advantages

After investigation, it was found that For example, medical image analysis requires the support of 3D neural network architecture, which is very different from the 2D image deep neural network common in the industry. Reporters from

learned that most teams participating in the Tianchi Competition use 3D image data solutions. The models themselves are not large, but the data input each time exceeds the memory capacity of commonly used graphics cards. In addition to 3D data in the medical industry, there is also MRI 4D data with time series, which are more dependent on memory. At this time, GPUs often face problems when processing 3D and 4D image data.

In order to make excellent medical AI solutions, it not only requires computing power in a narrow sense, but also has strong storage capacity to support it. Therefore, medical treatment poses new challenges to the underlying AI facilities.

Based on the above situation, Intel The advantages of Xeon Phi processor , which is specially designed for deep learning, begin to be highlighted. Just like a CPU, it can directly read the memory, and the memory capacity that a single processor can support and use is Up to 384GB.

Many participating teams described the Intel Xeon Phi processor as follows: "We must first understand the most important point, it is a CPU, not a card. You can understand it as a high-performance, capable CPU products that can do deep learning and can be directly connected to memory. "

Before this, thanks to its significant advantages in face recognition and common image recognition, GPU has become the first choice for most developers to train models.

Zhou Xiang, senior technical director of cloud computing at Intel China Data Center, revealed to Lei Feng.com that he often sees in colleges and universities that many medical image analysis models that students have worked so hard to design cannot run on GPUs.

To do this, the students had to reduce the pixels of the 3D lung image data, split it into multiple small blocks, and then identify them one by one. This "compromise" approach of

will cause two major problems: first, reducing pixels will lose detailed information; second, block recognition may produce errors. The

Tianchi competition champion team (Peking University LAB2112) also has a deep understanding of this. Captain Hu Zhiqiang said that lung CT image analysis has high storage requirements. The GPU platform is limited by video memory. In order to better analyze CT images, Analysis often requires multi-GPU collaboration or even multi-machine collaboration, but these projects are difficult to implement, especially for layers such as Batch Normalization that require multi-GPU synchronization.

In contrast, the Xeon Phi platform can utilize large-capacity memory, which can better meet the memory requirements of CT image analysis. At the same time, there is no synchronization problem between multiple devices. In terms of

software, the reporter learned that through the efforts of and Intel , most of the commonly used deep learning frameworks now have optimized versions for the CPU. But further, Intel has also deeply customized its own optimized version of the Caffe framework and algorithm library specifically to meet the needs of this medical AI competition. By adding and optimizing dozens of key operations such as three-dimensional convolution layers, three-dimensional deconvolution layers, and loss function layers, it has greatly ensured that the computing power of the Xeon Phi platform can be fully utilized, combined with the platform's own storage Its powerful advantages enable it to truly effectively support deep learning applications in three-dimensional medical imaging.For developers,

is relatively easier to develop on Xeon Phi processors. Although its bottom layer is deep learning technology, the upper instruction sets are all X86, and some libraries provided by and Intel can also be used. By covering the lower layers, developers can quickly become familiar with it as long as they are familiar with programming on traditional CPUs.

"Many people like to talk about how good their MicroBenchmark is on the GPU. Indeed, everyone is used to preprocessing after getting the image. But placing too much emphasis on MicroBenchmark often ignores the errors caused by preprocessing. So if you If you want to make a more practical finished product, you should focus on more macro tasks and focus on solving solution-level problems. The Intel platform will make the realization of the entire goal easier," Zhou Xiang said: "So Intel hopes to provide everyone with a new idea through the Tianchi Competition, so that they can understand that whether it is medical imaging or high-precision video analysis, your programming methods and training model methods are not affected by the memory bandwidth and capacity of existing hardware. "

• Secondly, the general deep learning open source framework can no longer meet the needs of medical AI applications. A more optimized framework is an essential tool.

Hardware platforms are like land, and if you want a good harvest on the land, you cannot do without the tools for farming. They are software frameworks.

In this competition, Intel customized 43 new features for the deep learning framework Caffe that surpass the open source version to support contestants’ model innovation. Due to the higher abstraction level of Caffe,

has significantly better performance than other frameworks, but it also suffers from poor flexibility. To this end, Intel has customized 43 new functions for its flexibility shortcomings to provide better support for medical image analysis problems.

At the same time, Intel also contributed 35,000 lines of framework code and 6,000 lines of reference model code to Tianchi Software to protect model training. The results showed that 80% of the problems encountered by the players of each team during the competition were verified by Intel in advance.

"The experience accumulated through this competition has laid a good foundation for the expansion of deep learning frameworks in the future. We will also consider introducing more frameworks in the future."

Xeon Phi platform compared to GPU in medical imaging AI applications There are so many advantages on the GPU platform, so how should developers who are accustomed to the GPU platform achieve smooth migration?

In response to this problem, the Tianchi champion Peking University team also said that this matter was troublesome at first, but it was successfully solved in the end.

"We first try to divide the entire process into different modules. This modular development method can make the code structure clearer, make each part functionally independent, facilitate troubleshooting, and avoid many problems; we implement the PyTorch version framework in the first stage At that time, the issue of the rematch had been taken into consideration, so as many data processing modules as possible were independent of the deep learning framework, and libraries such as NumPy were used to implement all operations.

In addition, the developer team should also spend more time reading the source. Code, because we used the Caffe version of Intel optimized for the CPU in the second round, some functions have been changed compared to the public version of Caffe. In order to better use these functions, it becomes very useful to read the source code. Necessary, and understanding the source code can also help us find errors faster when we encounter problems. For example, during the implementation process, we found that the network initialization method used at the beginning did not handle three-dimensional convolution, so we corrected it ourselves. The initialization method has improved the results," team leader Hu Zhiqiang told Lei Feng.com.

Hu Zhiqiang continued that the entire process, whether it was code migration or other major technical issues, was inseparable from the service and support of Intel technical experts.

"The most impressive time is when we use Intel optimized version of Caffe, we often encounter Segmentation Fault errors. The difficulty of troubleshooting this problem is that the error does not often recur, which means that sometimes when running the code Errors will happen and sometimes they won't, making it difficult to figure out the cause of the error.After reporting this problem to the Intel team one night, the staff quickly started to help us troubleshoot the problem and gave us a reply the next day. The reason for the error was that we used an in-place operation. Cause memory overflow. "

" Intel staff sacrificed their rest time, stayed up late to help us debug the code, and found out the problem in a short time, which also reflects the strong technical strength of their team. "

Now, this young Peking University team has cooperated with some medical institutions, including several affiliated hospitals of Peking University, and medical experts have also given very good evaluations of their system.

3. Intel Medical AI Vision

At the end of this interview, Zhou Xiang presented Intel 's vision for future medical AI from the perspective of the whole society and individuals:

People's understanding of Intel has always been that of a chip company. It seems that Intel is much more than just developing chips. As Intel CEO Krzanich said, we are a data-centric company.

There is a lot of data in the medical industry that has not been properly developed. We hope to join forces with excellent companies such as Alibaba and Lingkrypton Technology to revitalize data from all walks of life.

Intel especially hopes that through this medical AI competition, those who really do medical care, those who own data, and those who formulate policies. People can see that a medical model supported by artificial intelligence is more efficient, safer, and more valuable.

I think technology companies cannot choose to wait and see because of policy and regulatory issues, waiting for the implementation of policies and regulations. Action. On the contrary, our corporate mission is to consider whether we can accelerate the implementation of AI technology by supporting innovators around the world, so that more policymakers and regulatory agencies can clearly feel that our innovation results and actual use results are constantly improving. , thus indirectly promoting the implementation of policies and accelerating the recognition and recognition of artificial intelligence by the whole society.

From a personal perspective, as a small individual in a company and society, the thing that makes us technology practitioners feel the happiest. There is nothing better than long-term efforts that can not only benefit the world, but also help yourself and those around you