VBD 
As an artificial intelligent system combining the power of leading medical intellectuals and hundreds of thousands of medical images to complete five diagnostic modules, VinDr can now effectively support doctors in the diagnosis, detection and localization of various lesions on the lung, liver, breast, and spine. VinDr takes only a few seconds for each scan with average accuracy of over 90%.
According to GLOBOCAN statistics, it is estimated that in Vietnam in 2020, for every 100,000 people, 159 were diagnosed with cancer and 106 died of it. Thus, Vietnam has moved up 6 to 8 places on the world cancer rankings, respectively at number 91 and number 50 out of 185 countries in terms of incidence rate and mortality rate. Liver, lung, and breast cancer are among the most common diseases affecting both men and women in Vietnam.
Along with cancer, bone and joint diseases are also becoming a serious threat to the Vietnamese people’s health. Statistics show that the prevalence rate of degenerative spine disease in the general population is around 35%, and this number is up to 89% for the age group of 60 to 69.
To deal with two major issues outlined above, part of the solution resides in the development of the disease diagnostic and screening techniques. In addition, Vietnam is facing the problem of the overloaded healthcare system, as well as the unequal distribution of medical resources between upper and lower level hospitals. Therefore, increasing the quantity, quality, and speed of diagnosis is the most pressing issue right now.
VinDr and the journey to find a high-tech solution
From the urgency to solve three challenging problems which are screening for deadly diseases, diagnosing bone and joint illnesses, and reducing pressure for the upper-level medical system, a team of scientists and engineers from the Medical Image Processing Center of VinBigdata has researched and developed VinDr AI, which is a set of models that automatically diagnose and localize lesions on medical images. After one year since its publication, VinDr AI has completed five modules on chest X-ray, mammogram X-ray, spine X-ray, chest CT, liver CT diagnosis. This result has been achieved thanks to the combination of world-class technology, large-scale Vietnamese data, and the minds of the top doctors in Vietnam.
In terms of technology, VinDr is built on computer vision (CV), machine learning (ML), and deep learning (DL). In comparison with traditional methods, which are predefined algorithms, deep learning is outstanding in that it allows VinDr to self-learn pathological features based on large-scale data to output the fastest and most accurate diagnosis.
Also, VinDr is trained on large-scale clinical datasets (containing over 300,000 annotated medical images of various types such as X-ray, CT, MRI, etc.) collected from a variety of hospitals across the country. After patients’ identifying information is removed, the image data is preprocessed, then stored on the Label-PACS system for doctors to access and label remotely.
With the above technology and database, VinDr is capable of automatically diagnosing multiple studies simultaneously in real time. The average accuracy of the lesion localization task reaches over 90%. Furthermore, VinDr demonstrates its superiority in terms of speed, taking only a few seconds for each scan. It’s worth noting that the speed and accuracy of the system do not change overtime, and it can operate around the clock.
Thus, VinDr is possibly the solution to the current major difficulties of the healthcare system in Vietnam. To meet the growing demands in early detection of disease risks, besides training a new generation of medical staff of sufficient quality and quantity, AI systems like VinDr will be a solution that is quick, strong, and long-lasting enough to continuously address medical problems. When applied in hospitals, artificial intelligence will not entirely replace the role of radiologists, but provide an additional opinion for doctors to refer to. In other words, the system will serve as a useful support tool, like an assistant who works together with the doctor. This will noticeably improve the accuracy of the diagnosis.
Diagnosis of lung diseases
The main difficulty in the treatment of lung cancer is how to raise the number of patients who are diagnosed early. An infection that damages the airways and causes illnesses like bronchitis or other chronic infections is one of the unmistakable signs of lung cancer. Using a chest X-ray to locate the lesion, doctors can completely recognize a chronic lung infection at an early stage. As a result, chest X-ray is widely used to detect anomalies before moving on to further interventions like computed tomography (CT) or biopsy.
Based on this fact, the VinBigdata development team has designed and completed two related models which are VinDr-ChestXR (chest X-ray diagnosis) and VinDr-ChestCT (chest CT diagnosis).
VinDr-ChestXR was developed and evaluated on hundreds of thousands of chest X-rays collected from hospitals in Vietnam. With the support of AI, the system can currently detect six lung illnesses and localize 22 common abnormalities on chest X-ray images in under a second. During actual implementation at major hospitals in Vietnam which are 108 Hospital, Hanoi Medical University Hospital, Vinmec Times City Hospital, and five other hospitals in Phu Tho province, the evaluation results showed that at the 108 Hospital about 10.5 percent of diagnoses changed after doctors consulted AI, and the average consensus of doctors with AI was 90 percent. This result is similar to that at the Hanoi Medical University Hospital, with the respective rates of 4.8 percent and 89.5 percent. On average, the accuracy of VinDr-ChestXR in identifying lung illnesses was over 90 percent.
Besides chest X-ray diagnosis, VinDr also provides a chest CT diagnostic model (VinDr-ChestCT). Trained on 4000 CT scans and evaluated on over 1000 cases, the software is able to detect 5 lesions, including consolidation, pulmonary edema, mediastinal lymph nodes, pulmonary nodules/mass, and pleural effusion. In comparison to X-ray diagnosis, the diagnosis on computed tomography allows clear detection of the opacities, detailed evaluation of the size as well as the degree of lung damage. Currently, VinDr-ChestCT has been deployed in six hospitals and medical centers in Phu Tho, Binh Dinh, and Hai Phong. VinDr has proved to be 78% accurate and has a reading speed of 10 seconds per case. It is expected to become a widely used method in the early diagnosis, screening and treatment of lung cancer.
Diagnosis of liver diseases
In Vietnam, liver diseases, especially liver cancer, are currently the primary cause of death. The most popular method for reliable identification of liver lesions is liver tomography, which is more effective than the X-ray method. However, reading CT scans of the liver requires a lot of time, as well as the involvement of experienced doctors. As a result, VinDr-LiverCT was created as an AI tool to assist doctors in making faster and more accurate diagnoses.
For VinDr-LiverCT, the model was trained on 3000 CT cases and evaluated over 1000 cases. Data is gathered from various large hospitals in Vietnam, including 108 Hospital, Hanoi Medical University Hospital. The model, which is based on deep learning technology, can detect four lesions: mass, calcification, cystic lesion, and biliary dilatation in an average of 10 seconds per case and with an accuracy of 80.5 percent. Similar to VinDr-ChestCR, this technology is currently being implemented in six hospitals and health centers in three provinces, including Phu Tho, Binh Dinh, and Hai Phong.
It is expected that VinDr-LiverCT will be further developed to achieve the ability to identify 10 liver disorders, including different types of liver cancer, as well as localize 24 common liver abnormalities.
Diagnosis and screening of breast cancer on a large scale
According to the Food and Drug Administration (FDA) of the United States, mammography X-ray is the most common and least expensive technique of screening that reduces mortality from breast cancer. X-rays possibly show calcifications and masses, which are the early signs of cancer. Therefore, this method is utilized both in diagnosis and screening, to test on a broad scale when the patient has no specific symptoms or to observe the lesion more clearly.
Hence, in order to solve the breast cancer problem, the VinBigdata team has developed a model to diagnose mammograms X-rays (VinDr-Mammo). To be able to perform multi-task on mammograms X-ray, the model was trained on 50,000 mammograms collected at major hospitals across the country. Thanks to modern technology and large scale datasets, the model can classify the density of breast parenchyma, as well as localize many different types of lesions on mammograms with an average accuracy of over 85 percent and a diagnosis time of less than 05 seconds for each case.
Moreover, the model also features BIRADS capability, which is a mechanism for classifying mammogram data according to the available scale. Based on the BIRADS 1 to BIRADS 6 scale, VinDr-Mammo indicates whether the lesion occurs or not, and whether the lesion on the breast is benign or malignant. This is one of the first factors for the doctor to decide whether to conduct a biopsy in order to accurately identify signs of breast cancer.
In fact, the VinDr-Mammo’s test experiences at major domestic hospitals like 108 Hospital, Hanoi Medical University Hospital, Vinmec Times City Hospital and 05 hospitals in Phu Tho province has proved that on average, over 10 percent of the diagnoses change the results after the doctor consults the AI. Additionally, the average consensus of doctors with AI also reached 84 percent at Hanoi Medical University Hospital.
Diagnosis of spinal diseases
With the VinDr-SpineXR model, VinBigdata is currently the pioneer in applying AI to X-ray diagnostics of the spine. This model has the ability to classify photographic films (abnormal or non-abnormal) using a binary classifier. The training dataset of the model consists of DICOM formatted spine X-ray images collected from a variety of hospitals in Vietnam and labeled by experienced radiologists.
Developed based on deep learning techniques, VinDr-SpineXR can classify normal and abnormal spine X-ray images with 90 percent accuracy using AUC measurement (area under curve indicates the ability to distinguish patients with and without disease). The model accuracy in detecting and locating lesions reached mAP as 0.55 (mAP is the average accuracy index, allowing to measure the performance of the object detection models in the image).
Currently, VinDr-SpineXR is supporting doctors in the detection of six common kinds of spinal injuries in Vietnam, including bone spurs, disc stenosis, surgical materials, narrowing of the adapter hole, spondylolisthesis, and vertebral collapse. Highly appreciated for both quantity and quality of diagnosis, VinDr-SpineXR is effectively supporting spine radiologists at six domestic medical facilities including Phu Tho Provincial General Hospital, Phu Tho Obstetrics and Gynecology Hospital, Cam Khe, Thanh Ba, Thanh Thuy Medical Centers and Binh Dinh Provincial General Hospital. This is expected to be the solution to the problem of osteoarthritis, which is becoming more common, especially in younger people in Vietnam.
Along with five implemented models, it is expected that in the near future, VinDr will continue to improve two more models for diagnosing brain diseases, including cranial CT diagnosis (VinDr-BrainCT) and cranial MRI diagnosis (VinDr-BrainMRI). Also, currently VinBigdata Medical Image Processing Center is launching a new project, which is the application of AI in gastrointestinal endoscopy. The project aims to bring AI integrated with the endoscope. When the doctor moves the endoscope probe, the machine can see the lesions on the spot and give instant suggestions. This is a new direction in diagnostics for moving images, promising to be a step forward for the medical imaging industry in Vietnam. Combined with developed features and new breakthroughs, VinDr AI is expected to become a reliable medical imaging assistant for radiologists, contributing to improving the quality of medical examination and treatment, as well as improving community health.
