As soon as the AI pathologist came out, he refreshed the highest level of AI-assisted cervical cancer screening at home and abroad that has been published internationally!
On the basis that the negative rate is higher than 60%, the correct rate of negative interpretation of cervical smear is higher than 99%, and the detection rate of positive lesions is over 99.9%. Cytopathologists read cervical smear under microscope, which takes an average of 6 minutes per case, while AI identification takes only 36 seconds. The interpretation speed of AI-assisted cervical cancer screening model is 10 times that of manual interpretation!
The all-media reporter of Guangzhou Daily learned from the release conference of artificial intelligence-assisted cervical cancer screening in China Bio-industry Conference that Jinyu Medicine and Huawei technology have made a breakthrough together. Based on pathomorphology, the teams of both sides trained an accurate and efficient AI-assisted cervical cancer screening model with the diagnostic criteria of pathologists for the first time through deep learning technology.
Text/Guangzhou Daily All-Media Reporter He Xuehua Correspondent Zhang Jinju
Pathology is the gold standard of diagnosis, but there are too few pathologists.
Pathomorphological diagnosis is recognized by the medical community as the "gold standard" for disease diagnosis.
Dr. Liang Xiaoman, a well-known cytologist in China, has been engaged in clinical cytological diagnosis in the Cancer Hospital of Sun Yat-sen University of Medical Sciences for decades. He pointed out that the body has pathological changes. What exactly does it mean? Benign or malignant? What is the degree? Diagnosis depends on different morphological changes of pathology.
However, only 5,000 cells are qualified for a pathological glass slide, and there are tens of thousands of cells. If pathologists look at it one by one, it will be exhausting and very boring to look at 100 pieces a day. Therefore, a "pathological person" sees 90% negative films every day, and gets excited when he sees positive films. "It’s really not easy to see people get excited when they are sick," said Liang Xiaoman. This is a double high effort of brain power and physical strength.
There is a great demand for pathological diagnosis, but there is a shortage of pathologists.
Liang Xiaoman said, for example, like cervical cancer screening, cytology should be combined with HPV testing. In China, there are more than 350 million school-age women aged 25-65, and screening once every 3-5 years requires 70 million people to be screened every year. But in fact, since 2009, women’s screening for two cancers has been promoted, and the total number of cervical cancer screening in 10 years is only a little over 70 million, that is, the population coverage rate is only 21.4%.
Medical research shows that the population coverage rate of screening should reach 80% in order to reduce the incidence of diseases in the population.
The lack of disease screening ability lies in the long-term scarcity of pathologists in China. Reading films under traditional microscope needs to be based on human visual interpretation, knowledge accumulation, skills and talents, and the training period can often be as long as 10 years. The data shows that there are currently less than 20,000 registered pathologists in China, and the gap is still as high as 80,000-100,000 based on the standard of 1-2 pathologists per 100 beds.
AI-assisted cervical cancer screening made a major breakthrough
"So, I am very excited to hear that AI pathologists have made a breakthrough in research and development!" Liang Xiaoman said that it was initially seen that "this AI pathologist" can reduce the workload of manual interpretation by about half, which means that the other half of labor and energy can be devoted to improving the positive rate interpretation, and the chances of pathologists being complained of "seeing less, not seeing correctly and seeing slowly" will also be greatly reduced.
The AI pathologist mentioned by Liang Xiaoman refers to the AI-assisted cervical cancer screening model jointly developed by Jinyu Medical and Huawei Cloud.
There are about 500,000 new cases of cervical cancer in the world every year, and the incidence rate in China accounts for 26%. As long as the disease of cervical cancer is screened and treated early, its cure rate and five-year survival rate can be greatly improved.
According to reports, this development is based on 43.5 million cervical cytology screening samples over the years, and nearly 200,000 image blocks are selected for accurate labeling and AI-assisted screening model training.
First of all, it is marked by many pathologists to ensure that the model can fully learn the morphology of various pathological cells. Based on the block-level labeling of these images, the AI model first classifies the samples and efficiently distinguishes the samples with high positive cell density.
Then, the AI model further accurately identifies positive cells to ensure that accurate auxiliary interpretation results can still be obtained on samples with low positive cell density. In order to ensure the correct results, the algorithm also selects a series of suspicious local visual fields and submits them to the pathologist for final review.
At present, the AI model has been verified with the verification set of more than 20,000 sample data. The verification results show that the negative rate of the model is 61.9%, the accuracy of negative film interpretation is higher than 99%, and the detection rate of positive lesions is over 99.9%. This is the highest level of AI-assisted cervical cancer screening at home and abroad.
Application of results:
There are at least three to five years before landing.
The development of AI pathologists can be described as an important practice of digital medicine and precision medicine, which provides strong support for accurate diagnosis in the future. Once the results are applied, it is expected that the workload of single cytological examination by pathologists will be reduced by more than 60%, and the screening efficiency will be greatly improved. Usually, it takes 6 minutes for cytopathologists to read cervical smear under microscope. The AI ? ? recognition takes only 36 seconds, and the interpretation speed is 10 times that of manual interpretation.
At the same time, by comparing and verifying the results of AI tips, doctors can make manual interpretation more conveniently and accurately, which can bring effective help to the quality control of pathological examination and the training of pathological students.
In terms of greater social benefits, in the future, once AI-assisted cervical cancer screening is applied and promoted, it will greatly improve the coverage and service frequency of cervical cancer screening services, so that the screening quality of school-age women can approach the level of developed countries and promote early screening and early treatment of cervical cancer.
However, experts also admit that the results are at least three to five years away from landing.
First of all, we should continue to invest more sample data to verify the AI model on a large scale, and at the same time make more professional annotations to improve the model, so that its specificity will be continuously improved under the premise of ensuring sensitivity, and even adapt to and be close to the reagent consumables and production level of primary hospitals.
Secondly, it is ready to start the registration of CFDA. It is understood that there is no clear guiding process and cycle at present. "Many analysts in the industry believe that the large-scale application of medical AI such as pathological AI and imaging AI in China is estimated to take 3 to 5 years."
"AI pathology" vs "AI image" is more accurate.
In recent years, AI medical care has become more and more hot, which belongs to the field of auxiliary diagnosis. AI images have developed even faster and entered the registration application process earlier than AI pathology. So, which is stronger, AI pathology or AI image?
In this regard, some cytopathologists believe that AI pathology is "more difficult, more refined, more lacking and slower".
Because the image is based on black and white, density, size and edge, and the pathology is polychromatic and reaches the cellular level, it is not just an auxiliary diagnosis of "whether it is a tumor" or "whether it is a lesion".
In other words, AI imaging doctors can "allow" the conclusion that "this may be a cancer, a benign tumor or a malignant tumor", while AI pathologists can’t do this, but tell "what is this, is it a cancer, and what kind of cancer it is".
