With the continuous in-depth development of artificial intelligence technology, professional medical artificial intelligence platforms have gradually emerged. It is recommended to choose a professional integrated platform to save the work of platform construction and debugging and focus more on model training and system application. At the same time, the developed artificial intelligence system also has high reliability and high efficiency performance.
In the context of the future medical integration, the platform built should have the characteristics of weak coupling and strong compatibility to meet the compatibility and integration requirements between artificial intelligence systems, medical equipment and hospital information systems, and improve the performance of medical artificial intelligence systems.
Based on the establishment of a professional medical artificial intelligence platform in the hospital, it works closely with the clinical departments of the hospital to select suitable disease types for the development of its diagnosis and treatment system, thereby improving the effect of diagnosis and treatment.
Start with the development and application of medical imaging artificial intelligence systems, and on this basis, further integrate more types of data such as medical record data, inspection data, patient daily health monitoring data, etc., so as to build a richer and more comprehensive medical big data , Lay a solid foundation for the development of a richer artificial intelligence system.
01 Construction of medical artificial intelligence platform
The medical artificial intelligence platform includes data resource layer, artificial intelligence platform and medical application layer.
(1) The data resource layer provides basic data. By collecting medical imaging data and medical record data from various departments, it breaks through the data barriers between business systems and provides a data foundation for the artificial intelligence platform.
(2) The artificial intelligence platform consists of computing capabilities, open source frameworks, algorithms and technologies. Computing power provides guarantee for the computing speed of the artificial intelligence platform.
Take the medical imaging data of lung nodules as an example. Each patient has an average of 20-30 films. Computer vision models commonly used in automatically identifying lung nodules, such as residual neural networks, can make dozens or even hundreds of layers of images. Neural network training is possible. Most of the companies that provide medical solutions for computing energy solution providers are companies with comprehensive medical informationization experience. These companies have been in the medical industry for many years and are familiar with the business processes of the medical industry. It has natural advantages in the development of artificial intelligence medical imaging.
These companies have a deep understanding of the application scenarios of hospitals and the demands of doctors, and can quickly combine artificial intelligence technology with needs to form products that meet the needs of doctors. In addition, these companies have a relatively wide range of customer channels and stable cooperative hospitals, making it easier to promote and implement artificial intelligence systems in hospitals.
At present, part of the application of artificial intelligence in the medical industry is applied to the medical instrument side with embedded systems, that is, artificial intelligence technology is used on the medical equipment side to optimize the performance of the equipment.
For example, use motion capture technology to judge the patient’s recovery, provide visual data and image display, and provide strong data support for doctors to formulate rehabilitation plans; the other part is based on image data and medical records in the data center to assist in image diagnosis and clinical decision-making And other fields.
Based on their original business, these companies have accumulated years of technology as their own advantages, and they have expanded into new business areas and new areas of medical artificial intelligence.
Deep learning for medical data providers is particularly suitable for applications with large amounts of data, such as large amounts of data generated by routine examinations.
The ability to improve the efficiency and accuracy of diagnosis is essential for the early diagnosis and treatment of diseases. It can be very useful in areas where the evaluation of images and pathological slices takes a long time due to the shortage of doctors.
As medical imaging providers, primary hospitals, specialist doctors, provincial hospitals and emerging independent imaging centers have an urgent need for artificial intelligence-assisted imaging diagnosis systems.
China’s medical imaging data is in the transition stage from traditional film to electronic film, and the signal-to-noise ratio of image data is relatively low. Even if doctors undergo long-term professional training, diagnosis conclusions are often limited by the doctor’s own experience, fatigue, and patience. .
Deep learning uses unsupervised or semi-supervised feature learning and hierarchical feature extraction efficient algorithms to replace manual feature acquisition. Although there are some irresistible factors, such as large differences in data quality, the above-mentioned human factors are reduced to a certain extent. The diagnosis is inconsistent, reducing the misdiagnosis rate.
The construction of medical artificial intelligence platform 15 The construction of medical artificial intelligence platform assists medical institutions to improve service levels, balance medical resources, and ease the pressure of seeking medical treatment, especially in areas where medical resources are scarce.
Medical institutions choose different construction models according to their own informatization level to help improve their own medical service level. Very high requirements are put forward. The huge amount of data makes the computing time of the computer become long. Therefore, building a supercomputing platform can not only shorten the computing time, but also improve the efficiency of medical treatment and reduce the waiting time of patients. This is in clinical applications is crucial.
Platform model 1: Building an independent medical artificial intelligence platform
The hospital uses a large amount of medical data to build an artificial intelligence medical platform independent of business systems, integrates multi-source heterogeneous data scattered in various business systems, and uses natural language processing technology to transform clinical description information into structured language to generate medical knowledge Atlas, valuable medical knowledge and treatment experience are retained and quickly copied to places with insufficient medical resources.
The construction period of an independent medical platform is relatively long, and there are many business systems involved in docking, and it faces more challenges in the construction process.
In order to obtain a better algorithm model, it is usually necessary to label medical data.
Even if unsupervised learning or semi-supervised learning is used, it is also necessary to input labeled medical data for model training in the early stage. Data labeling takes a long time, the threshold is high, and there are high requirements for labeling personnel. At present, the personnel engaged in data labeling are mainly experienced professional doctors, and the entire process is done manually.
At the same time, the awareness of collaboration among medical system IT vendors needs to be further improved. As the “blood” of medical development, data needs to flow freely among various systems. Breaking through the barriers between various business systems of hospitals is the key to the development of medical artificial intelligence systems.
Platform mode 2: Building an embedded medical artificial intelligence platform
As the business system supporting the normal operation of the hospital, the original information system of the hospital has a complex structure and huge transformation cost. The emerging artificial intelligence medical diagnosis system on the market can hardly replace the original business system.
In most cases, the artificial intelligence system provides a service interface, which is connected to the original business system, and the artificial intelligence technology is organically combined with the original business system.
Taking medical imaging as an example, the result output of a suspected lesion does not require the doctor to open another system, but the original image archiving and communication system (PACS) prompts the information of the suspected lesion.
This built-in artificial intelligence module can reduce the cost of system development. More importantly, this model does not change the doctor’s original diagnosis process and operating habits, and can reduce the learning cost of medical staff. The artificial intelligence system that does not change the established model is easier to be accepted by the hospital, and the utilization rate of the system is higher.
The embedded artificial intelligence platform does not rely on the data of the original system. Now that the importance of data is becoming increasingly prominent, there is no need to open the database of the original system, which can not only ensure the data security of the original medical system, but also increase the cooperation among various manufacturers, which is conducive to the promotion of artificial intelligence technology in the medical industry.
02 Three key elements of the establishment and application of medical artificial intelligence systems
The establishment and application of medical artificial intelligence systems need to deal with the following three key elements, and overcome the challenges in dealing with the three elements in order to succeed. The three elements are as follows: data, platform computing power, and deep learning algorithm model.
1. Data
The medical artificial intelligence system needs medical big data as the foundation, and forms a certain degree of intelligence through machine learning and other technologies to provide auxiliary diagnosis and auxiliary treatment functions.
Medical big data mainly includes medical textbooks, medical records, especially medical records for certain types of diseases, digital medical images, academic papers, etc.
For medical imaging artificial intelligence systems, digital imaging data, including CT, MRI, ultrasound, pathology and other imaging data, are needed as raw materials for machine learning.
Because medical record data, digital medical image data, etc. belong to the hospital’s intellectual property, the intellectual property ownership principles and management methods of artificial intelligence systems need to be continuously explored in practice.
There are many types of medical data, a wide range of sources, and data formats vary widely. Therefore, rapid data collection, integration, and processing are used to ensure the training and learning of artificial intelligence models. This is the basic challenge that needs to be overcome to develop artificial intelligence systems.
At present, when imaging artificial intelligence-assisted diagnosis systems are used in hospitals, they usually need to use the hospital’s imaging data to relearn and challenge model parameters to meet the needs of the hospital.
This is because in the key factor of image data, the standards used in image generation are inconsistent among hospitals.
For example, the standard of the dosage of the developer and the inconsistency of the equipment parameter settings cause the difference of the image gray level, etc., resulting in different image data for the same patient between hospitals. When used to support machine learning, the model parameters will also be different.
In order to be able to increase the applicability of artificial intelligence systems, it is necessary to quickly integrate data from multiple sources when developing artificial intelligence systems, so as to train more accurate and widely applicable artificial intelligence systems.
2. Deep learning algorithm model
In addition to processing data, the selection or development of deep learning model algorithms is also a major challenge in the development process.
At present, there are many deep learning algorithms, but these algorithms are difficult to directly apply, but need to be improved and developed, and then applied to data training, and continuous improvement and perfection in the training, in order to make the algorithm model more and more accurate.
Therefore, choosing appropriate algorithms or developing algorithms, and establishing a platform system for algorithm adjustment and improvement are one of the elements for the success of artificial intelligence systems.
Because the AI system is in its infancy, the model algorithm of the artificial intelligence system currently used in hospitals still does not completely meet the actual needs, and it needs continuous improvement. The improvement of the algorithm model is also an important work that continuously pushes the AI system to be more accurate. As shown in the figure below, according to the survey, the AI systems currently used in hospitals all need different degrees of improvement or upgrade algorithms.
3. Computing power of artificial intelligence platform
Building a powerful computing platform is one of the fundamental elements for the success of artificial intelligence development. Because deep learning requires a huge amount of data to be input to the training model, and the training model requires huge-scale operations to train the model to be intelligent, the computing power (computing power) of the artificial intelligence platform is a key element of its success .
At present, artificial intelligence computing platforms mainly use GPU chips, and medical imaging artificial intelligence systems rely on GPUs for training and learning. There are also some AI systems that use CPU, FPGA, high-performance processor (TPU) and other chips.
At present, major server manufacturers have also developed servers for machine learning and running artificial intelligence systems, such as Dell, New H3C, Lenovo, Inspur and other server manufacturers. NVIDIA has also developed the supercomputer DGX for the development and operation of artificial intelligence systems.
Most computing systems of artificial intelligence platforms currently use open source systems, and customized development is made on open source systems to meet the needs of their own products.
The mainstream open source systems currently in use include TensorFlow, Distributed Machine Learning Toolkit (DMTK), Caffe, etc.
Customized development on an open source platform requires very strong development capabilities, and the technical level of the development team is very high, because the development level determines the computing power and computing efficiency of the computing platform, and determines the accuracy of the artificial intelligence system.
Clara, a professional computing platform launched by NVIDIA, is well packaged and integrated with the computing power of NVIDIA GPU, and integrates a variety of machine learning models. It can provide professional support for deep learning and manual system operation, as well as processing image data and machine Learn to provide professional tools.
03 Mode comparison: Independently build medical artificial intelligence platform and embedded medical artificial intelligence platform
The development of medical artificial intelligence platform largely depends on the original degree of hospital informatization.
The foundation of the development of artificial intelligence is data. The hospital needs a large amount of historical data to support the scientific research work, medical history analysis, and treatment plan formulation of hospital doctors.
The willingness to cooperate and cooperate of medical system IT vendors affects the application of artificial intelligence technology in medical institutions. As the “blood” of medical development, data needs to flow freely among various systems. Breaking through the barriers between various business systems of hospitals is the key to the development of medical artificial intelligence systems.
The medical field puts forward higher requirements on artificial intelligence technology. Medicine is a systematic and complete system. At present, many researches of artificial intelligence companies in the medical field are focused on the identification of a single disease. This is valuable for academic research, but the artificial intelligence-assisted diagnosis of a single disease alone is useful for actual clinical practice. The work is of little significance. Medical institutions say that the recognition of a single disease has limited appeal to them.
Artificial intelligence technology needs to meet the basic clinical applications, support the identification of most diseases of a certain organ or support the identification of a series of diseases, before it can generate commercial value, so as to further promote related research and produce sustainable economic benefits. At the same time, the design of the product needs to conform to the doctor’s daily operating habits and diagnostic procedures.
Ultrasound detection means that doctors make a diagnosis when they see real-time images during the operation. This requires artificial intelligence technology to support real-time diagnosis, and there is a higher demand for computing power. If it is recognized according to the traditional collection first, it violates the doctor’s operating habits and diagnosis process. Therefore, the development of artificial intelligence technology in the medical industry not only depends on the development of technology, but also requires talents with in-depth understanding of the industry.
