Breast mammography AI landing clinical: 16 million pixel second reading, level close to senior doctor

Breasts have caused countless wars and praises, but they are also the most popular for cancer, such as breast cancer.

According to statistics from the National Cancer Center of China, breast cancer ranks first among all the malignant tumors that women may suffer, accounting for 16.51%. Worldwide, breast cancer ranks first among women in both developed and developing countries.

In order to detect breast cancer early, the mammography mammography is used in clinical practice. This method is mature, simple, inexpensive, and reliable. Multiple medical guidelines suggest that high-standard mammography screening and review can detect most preclinical stages of breast cancer, effectively reducing breast cancer mortality, and reducing unnecessary disability or avoiding traumatic treatment.

But in China, whoever reads the film has become the number one problem. Mammography X-ray molybdenum target reading is difficult and the culture period is long, which makes the doctors who are specialized in breast X-ray molybdenum target reading extremely scarce. There are only a few hundred people in the country; at the same time, due to the difference in reading levels between different levels of doctors Huge, the consistency of reading is also difficult to guarantee.

To solve these problems, recently, a domestic AI company offers another solution ideas --AI interpretation according to the medical chart.

It is reported that the mammography X-ray intelligent diagnosis system is developed and launched in accordance with the years of medical treatment. The AI ​​system relies on powerful algorithm innovations and real breast image data of thousands of hospitals in the top three hospitals to achieve a second-level reading of mammography images, with glandular classification, lesion detection, sign description, Intelligent BI-RADS typing and other functions, and can automatically generate structured reports for use by imaging physicians.

In the clinical practice of several top three hospitals, the system demonstrated strong lesion detection and learning ability. It can not only detect breast lumps, calcification, structural distortion and asymmetry, but also achieve the risk classification of cancer lesions based on the detection of the whole lesion type, and assist doctors to identify high-risk lesions.

The breast is beautiful, but it is not easy to see the internal lesions clearly.

Different from the fact that hundreds of images can be generated by layer-by-layer scanning, the lung CT of the three-dimensional structure of the lungs, the physiological structure of the breast tissue, and the principle of X-ray vertical irradiation are restored, so that the conventional examination position of the bilateral mammography target is only There are 4 single pieces in the MLO position and the CC position 2 individual positions. Therefore, only the compression plate can be used to make the female breast as thin as possible, so that the internal tissue of the breast can be analyzed and analyzed, in order to capture the image of the molybdenum target as clear as possible, and then to find the location of the lesion.

However, this imaging method must overcome gland obscuration and structural noise. Just like a hunter shuttles through the forest floor where the light spots are spotted, the position of the prey on the top of the tree is judged only by the change of the spot and shadow on the ground. Therefore, the reduction of 2D image results to 3D breast tissue poses extremely high demands on the professional ability of the reading physician.

Compared with the more obvious calcifications and masses, the structural distortion and asymmetry of about 30% of the total number of lesions are difficult to detect, especially when the doctor is not experienced or tired, it is easy to miss the diagnosis.

Professor Peng Weijun, director of the Breast Imaging Group of the Chinese Medical Association Radiology Branch and director of the Department of Radiology, Fudan University Cancer Hospital, has more than 30 years of experience in breast X-ray mammography. He said that he wants to achieve the "senior" level. Must have a solid anatomical foundation and diagnostic imaging skills, rich spatial imagination, sufficient clinical experience, several excellent instructors, and a 5-10 year growth cycle.

The high demand for talents makes the elderly mammography experts extremely scarce. Looking at the Mammography Group of the Imaging Group of the Chinese Medical Doctor Association, there are only more than 100 experienced professional mammography mammography readers, and there are only 50 "expert-level" people. Scarcity, these experts have to meet the reading needs of hundreds of millions of women with mammography in China.

"The signs of breast lesions are not as typical as those of pulmonary nodules. There are large differences in the diagnosis of many lesions. For example, structural distortions, some experts believe that the structure is distorted, and some doctors think it is not. Plus Asian women are mostly dense. Sexual breasts increase the probability of inconsistent readings.” Professor Peng Weijun revealed, “The same mammogram results, an experienced senior doctor and a low-grade doctor can make a difference of 30%. Even higher. Therefore, the potential of artificial intelligence applications in this field is huge."

The difficulty of developing a molybdenum target AI is beyond the reach of ordinary companies.

“From the beginning of R&D, we followed the latest ACR guidelines, and also referred to the NCCN guidelines, ACS guidelines and the latest consensus on breast cancer diagnosis and treatment in China. Whether research and development or engineering and technical personnel, we must start from the beginning to learn breast image and breast. Cancer-related knowledge, and long-term work with clinicians, in-depth understanding of the workflow and AI application scenarios, understanding the pain points of doctors." According to Lin Qiang, director of medical products products.

The professionally labeled clinical data is not only the key to building an AI model, but also the cornerstone of this breast AI.

"This AI brings together thousands of breast cancer imaging data from top three top hospitals across the country. It has complete posture, advanced equipment, professional shooting and clear images. It is the top breast tumor database in China." Lin Qiang is quite confident. Said, "In the annotation, everything from hardware to software has been carefully designed."

In order to make the description of the lesion more comprehensive, the R&D team “lives” in the clinical line, consult experts, read the guide, and write detailed labeling rules for each symptom description.

In order to make the doctors see more clearly, the R&D team has prepared a professional 5M professional reading screen to make the viewing more clear and less visual pressure.

In order to alleviate the pressure of labeling physicians, instead of simply pursuing the progress of labeling, the R&D team recruited dozens of highly qualified professional physicians to form a labeling team to spread the pressure and avoid the phenomenon of grabbing progress.

To ensure the quality of the annotations, each mammography image is labeled with at least 5 doctors. Only the results with a highly consistent result are recognized. The controversial label will be judged by a higher-ranking doctor and submitted to the team for review and voting. Finally, the authoritative expert will give the result.

In order to more effectively supervise the labeling process, the R&D team even developed a special label management system before building the AI ​​model.

High data collation, associated costs, cumbersome labeling process, and numerous controversial points have caused the R&D team to collapse, and the final model results have not lived up to this effort and expectation. In the process of landing hospitals, the system Constantly highly praised by experts.

“Our doctors in Europe and America see the molybdenum target of 10 patients a day, but in China, this number is at least 50. Associate professors who need to sign the film, 100 people or even 150 people a day. It is a common practice. Time is tight, the task is heavy, and no lesions can be missed. The physical strength and spirit of the doctor are under high pressure for a long time.” Professor Peng Weijun revealed that “the artificial intelligence system can greatly improve the speed and accuracy of the detection of the lesions and reduce errors. Inspecting the phenomenon of leak detection, while improving the consistency of reading, the doctors are freed from heavy mechanical labor and engaged in truly innovative work."

With the increase of international academic exchanges and the return of more and more top AI experts and scholars, it is not uncommon in the industry to use overseas public data sets for medical AI research and development. Objectively speaking, the emergence of public data sets and generalized AI models has greatly promoted the development of medical artificial intelligence in China. However, in the development of mammography X-ray mammography, the AI ​​model based on overseas public data sets has encountered Waterloo.

Lin Qiang revealed that unlike the breasts in Europe and America, the mammary glands of Chinese women are mostly compact, glandular obscuration and structural noise are more obvious, and the normal breast tissue and lesions are less differentiated, which puts forward the performance of AI system. Higher requirements.

“The mammography target AI developed on the basis of public datasets can be 95% or more than 99% sensitive in the laboratory. However, once clinical sensitivity is reached, there will be a serious decline, which requires long-term training and Data feeding, which invisibly increases the burden on clinicians." Lin Qiang said, "In addition, public data sets also have problems such as low image quality, poor labeling quality, and inconsistent labeling standards."

Lin Qiang takes a single MLO image in a mammography as an example. Under normal circumstances, the resolution of a molybdenum target image is as high as 4000x4000, and the total pixel is more than 16 million. In order to facilitate the publication of the public data set, it is compressed into the ordinary jpg format, and most of the pixels are lost. The performance of the lesion is greatly reduced, and the tiny lesions may even disappear directly. The level of the trained AI model can be imagined. know.

Therefore, the AI ​​advantage based on the real breast image data of Chinese women is undoubtedly revealed.

“We optimized the image reading algorithm for this AI system, enabling direct reading and second processing of 16 megapixel images without any catastrophic or crashing, ensuring that no matter how small the lesion is, it is clear. Without losing any details, the shape of the lesion is accurately restored. From this point of view, AI is far superior to humans." Lin Qiang said.

At present, most of the mammography examinations in China are concentrated in large and medium-sized cities. In the vast number of primary medical institutions, nearly 800 million urban and rural residents have access to medical services far less than cities, and mammography is no exception.

"In the future, when the breast AI is more mature, the cost of universal screening for breast cancer will be greatly reduced. The early screening methods for breast cancer will also change, from high-risk groups to all women of the appropriate age, thus greatly improving the early stage. Discovering the possibility of breast cancer and reducing the overall medical expenses of the society.” Lin Qiang said, “At the same time, by instrumenting the expert diagnosis and treatment capabilities and empowering primary care through AI, it will help alleviate the shortage of reading doctors in primary medical institutions. Difficulties, improve the level of early screening of breast diseases in primary health care institutions.

Animal Extract

Animal Extract,Natural Animal Extract,Nutrition Supplement Animal Extract,Edible Animal Extract

Allied Extracts Solutions , https://www.nballiedbiosolutions.com