Artificial intelligence spawning device inspection iteration upgrade
In a dozen square meters of office, a row of laptops is placed on a desk, and a group of young people are concentrating on operating the computers. This is a reporter who has recently entered the China Academy of Food and Drug Testing (hereinafter referred to as the China Institute of Inspection and Quarantine). The medical device inspections were performed during the photoelectromechanical room artificial intelligence (AI) team office. The current team, which is only 30 years old, is creating the “AI Miracle†for Chinese drug testing.
Pride: issued the first test report
In the past two years, the AI ​​deep learning algorithm has been rapidly applied in the field of medical devices, bringing new challenges to the inspection and inspection work of the Chinese National Laboratory.
It is reported that in 2016 the Chinese Academy of Sciences began to follow up on the latest developments in AI in the medical field. In the second half of 2017, when the number of AI products was increasing explosively, in order to meet the relevant inspection needs, the Chinese Institute of Inspection and Quarantine quickly conducted research on the inspection methods and specifications of AI products.
Earlier this year, the Chinese Academy of Inspections initiated the construction of a fundus image standard test database and a lung image standard test database. The former was completed on March 26 and formed product specifications and inspection specifications. At present, the Chinese Academy of Inspection and Testing is intensively testing the products of 11 companies, most of which have completed the tests of conventional GB∕T25000.51-2016, special requirements for mobile medical devices, and special requirements for network security, and are conducting tests based on the standard library. Offline performance testing.
On April 30th, the Central Inspection and Quarantine Bureau issued the first report on the detection of artificial eye imaging diopter retinopathy screening software. China's AI medical device inspection and testing has taken an important step.
Challenge: The requirements for building a database are extremely "rigorous"
According to Ren Haiping, director of the Department of Opto-Mechatronics, at present, the total number of image annotation images for the fundus image standard inspection database is 6327. All image annotations are based on expert consensus and are traceable. Sensitivity and specificity tolerances are <2% to meet the daily testing needs.
In the construction of the lung image standard test database, 611 lung image calibration tasks with nearly 10,000 nodule images were completed on June 3, covering the needs of the current stage of lung nodules AI medical device testing in China. Baked out.
“Building a database is a very difficult process.†Meng Xiangfeng and Wang Hao, researchers of the AI ​​team in the Department of Opto-Mechatronics, frankly stated that each of the three steps in image collection, image labeling, and data management is very challenging.
"The selection of calibration experts is very strict." Wang Hao said that taking the lung nodule image calibration expert as an example, 220 specialists from 112 top-three hospitals from 24 provinces registered and 185 experts were selected for online Plot the test, and then evaluate the accuracy, stability, and consistency of its annotation. Through the recommendation and examination procedures, 24 labelling experts and 15 arbitration experts were identified, of which 56.4% were experts with sub-higher titles.
Regarding data quality, the AI ​​team was unequivocal: three rounds of desensitization and data cleansing were performed on the lung nodule calibration data, and the organization developed online examination software, on-site nodule inspection software, and on-site nodule size measurement software. , analysis and evaluation software, and conduct multiple rounds of self-test and doctor pre-test. In order to ensure the observation effect and data security, special medical image displays were specially prepared for doctors, and local local area networks were set up.
Vision: Achieve multi-dimensional comprehensive evaluation
By using neural networks and optimization algorithms to learn and generalize the data characteristics of the training set, AI medical devices can achieve functions such as lesion detection, classification, image segmentation, signal analysis, and risk warning. The establishment of the two databases is conducive to a more accurate, more reliable, and more efficient diagnostic tool for the company's listing of clinical fundus glycocalyx lesions, lung nodule screening, and detection.
At present, the AI ​​team is concentrating on promoting the establishment of a lung image standard detection database. It will soon confirm the closure of the annotated image and conduct statistical analysis methods and assessment methods. It is expected that the confirmation of testing specifications can be completed before the end of this month, and a test report will be issued as soon as possible.
Of course, these are the immediate goals of the AI ​​team. According to Ren Haiping, in the future, the AI ​​team will have more work to do. According to the trends of AI technology and the new features of new products, the “testing and testing†of the quality characteristics and risks of AI medical devices will continue to be “iterativeâ€. While accelerating the establishment of a perfect standard database, we will adopt more plentiful means such as confrontation testing and physical modeling to realize scientific, comprehensive, accurate and rapid evaluation of new AI and new risks. In addition, more work will be done on the establishment and improvement of the quality specification of the open training data set to help the rapid development of AI companies. (Reporter An Huijuan)
Hydroponics is a subset of hydroculture, the method of growing plants without soil, using mineral nutrient solutions in a water solvent. Terrestrial plants may be grown with only their roots exposed to the mineral solution, or the roots may be supported by an inert medium, such as perlite or gravel. The nutrients in hydroponics can come from an array of different sources; these can include but are not limited to waste from fish waste, duck manure, or normal nutrients.
Greenhouse Hydroponic System,Greenhouse A Hydroponics,Greenhouse Flat Hydroponics,Greenhouse Vertical Hydroponics
JIANGSU SKYPLAN GREENHOUSE TECHNOLOGY CO.,LTD , https://www.engreenhouse.com