The latest study of Ali: Relying on AI to identify ordinary CTs can detect early esophageal cancer

Esophageal cancer is one of the high incidence of cancer in my country. Early, it can significantly increase the survival rate of patients, but the clinical medical community has always lacked good early screening methods. The Alibaba Barma Hospital joined forces to combine AI with ordinary flat -sweeping CT for the first time. It can effectively identify early esophageal cancer with a sensitivity of 93%and 98%of its specific degree. The accuracy has exceeded the doctor level. Research papers have been included in the Miccai 2022 of the International Faculty International Society held this week. At present, multi -central large -scale clinical verification is currently underway. It is expected to be widely landed in the medical examination items in the future.

Global Cancer Statistics 2020 (GLOBOCAN 2020) shows that the number of new cases worldwide worldwide in esophageal cancer exceeded 600,000, and the number of deaths exceeded 370,000, of which half of them came from China. The mortality rate of esophageal cancer is high because it is found that it is too late. After the treatment of early esophageal cancer patients, the 5 -year survival rate is good after treatment. Therefore, the State Health and Health Commission's "Guide to Diagnosis and Treatment of the esophagus (2022)" proposed the strategy of early diagnosis and early treatment.

Although there are many means of esophageal cancer, the accuracy is relatively limited, and it is easy to missed inspection, especially the lack of low thresholds and high reliability methods. Taking the gastroscopy biopsy as an example, it needs to extend the endoscope from the mouth into the digestive tract. The invasive operation causes the user experience to be poor and difficult to promote. CT image screening depends very much on doctors' superb discrimination, especially early esophageal cancer lesions, most of them are about 1-2 cm in diameter, and are similar to the muscle tissue of the esophagus, which is very difficult to distinguish.

The Dharma Hospital Medical AI team combined with the First Affiliated Hospital of Zhejiang University Medical College, Sun Yat -sen University Cancer Prevention Center, Shengjing Hospital of China Medical University, Sichuan Cancer Hospital and other institutions. The basic idea is to allow patients to do a flat -sweep CT examination with lower thresholds, and then use AI technology to identify whether there are esophageal tumors, benign or malignant. The flat scan CT is the most common CT scan. It does not require additional injection of iodine -containing film, which has been widely used for various medical examinations.

In response to the characteristics of the early esophageal cancer size and difficult to distinguish between normal tissues, the research team focused on improving the division algorithm model, introducing a global attention mechanism, and combining positions with positions to allow the segmentation model not only to focus on the part, but also the whole. In other words, AI needs to learn the shape of the esophagus, texture and other global characteristics to determine whether there are abnormalities such as asymmetric orientation or squeezing the esophagus wall. At the same time, local image details are also analyzed. , Greatly improved the ability of AI to identify early esophageal cancer.

The papers of the Miccai 2022 "Effective Opportonistic Esophageal Screening USINTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRASTRATTRASTRATTRASTRATTERATTERATTERAT. In independent testing, the sensitivity of the AI ​​recognition was 93%combined with AI identification (the identification of esophageal tumors was tumor), and the specific degree of specificity reached 98%(the identification of healthy esophageal was health). The research team also invited four doctors from radiology and radiotherapy in different hospitals to participate in the test. The experience ranges from 5 years and 14 years. The results of doctors' identification were weaker than AI. The research team also compares AI recognition with other methods reported by literature. The accuracy of AI recognition is basically equivalent to the quantitative results of gastroscopy, which exceeds blood tests and cytology.

AI recognition accuracy exceeds four doctors

"After the dissertation is submitted, we are still continuously deepening research. At present, more than 2,000 training cases have been added, and the accuracy has been further improved. In the independent test of more than 5,000 people, the sensitivity has reached 98%, and the specificity has reached 99.5%." The person in charge of the Dharma Medical AI team and IEEE FELLOW Lu Le said that this technology has opened the API call interface on the public cloud for doctors to experience and use it; in the future, it is expected to be used for conventional medical examination items to reduce the screening threshold for screening of esophageal cancer. Realize early diagnosis and early treatment.

Early esophageal cancer and benign tumors identified by AI, doctors fail to identify

MICCAI 2022 September 18-22 this year was held in Singapore. This is the world's most influential medical imaging top. The Dharma Medical AI team has a total of 4 papers. In addition to esophageal cancer, it is also colorectal cancer and skin. Breakthroughs were made in image analysis such as cancer, lymph detection. At present, the team is developing the AI ​​imaging system of precision cancer diagnosis and treatment, including cancer diagnosis and treatment technologies such as scale screening, accurate diagnosis, prognosis treatment, and response evaluation, covering multiple important diseases. The Dharma Medical AI team has long been committed to research in medical imaging and other directions. In the early stage of the epidemic, the CT image new crown pneumonia AI auxiliary diagnostic system was developed, and the Ministry of Science and Technology was rated as the national advanced collective of science and technology resistance.

API call interface: https://help.aliyun.com/document_detail/452777.html

Thesis link: https://link.springer.com/chapter/10.1007/978-3-031-16437-8_33

- END -