Engineer Nguyen Ba Dung, Medical Imaging Center, VinBigData (under Vingroup), has just won the first prize in a Kaggle competition titled SIIM-FISABIO-RSNA COVID-19 Detection.
The method is expected to be the solution to large-scale Covid-19 screening, thus addressing the urgent situation when the number of infections is increasing worldwide.
The only Vietnamese engineer to rank first at the global AI competition
The Covid-19 AI Detection Competition (SIIM-FISABIO-RSNA Covid-19 Detection) was organized by the Society for Imaging Informatics in Medicine (SIIM), The Foundation for the Promotion of Health and Biomedical Research of Valencia Region (FISABIO) and Radiological Society of North America (RSNA) from May 2021.
In this competition, participants need to develop an AI model to classify a chest X-ray as negative or typical, atypical, and unspecified pneumonia. For non-negative cases, the AI model have to localize areas of ground-glass opacity, then detect lesions and provide suggestions to help radiologists make fast and precise decisions, thus improving patient care by enabling the right treatment before the most severe effects of the virus take hold.
The winning model will be deployed into clinical applications to improve healthcare quality and efficiency on a global scale. Meanwhile, the source code will also be publicly shared for the world’s scientific community to continue to develop smart medical solutions.
“Currently, COVID-19 can be diagnosed via polymerase chain reaction (PCR) to detect genetic material from the virus or chest radiograph. PCR is highly accurate but takes from a few hours to a few days to produce results. On the contrary, chest X-ray can be obtained in minutes at a much cheaper cost than a test kit,” engineer Nguyen Ba Dung (Medical Imaging Center – VinBigData, Vingroup) shared about his desire to contribute to the fight against the pandemic while participating the competition.
Mr. Ba Dung said, to solve the above problem, he used a combination of results getting from 40 individual models, all of which are currently the most popular and best AI architectures such as EfficientNet, ResNet, YOLO, Faster R. -CNN, EfficientDet, etc. To ensure high accuracy in lesion detection and classification, the AI model is trained by world famous and large-scale chest X-ray datasets such as the CheXpert (Stanford University (UK)), the X-ray datasets of NIH (US) and the RNSA, in addition to the 6,000 images the competition provided.
Surpassing more than 1,300 teams with nearly 1,800 multinational competitors from world’s well-known technology corporations such as Google, Nvidia, Aillis, Kyoto University (Japan), engineer Nguyen Ba Dung won the top prize of $30,000.
AI application in disease diagnosis
Going beyond the competition result, Ba Dung’s achievements have practical significance in reducing cost and workload for the current testing system. With computing power, the AI system can predict millions of infections within just a few minutes and can be considered the most optimal solution for large-scale disease screening.
This is also the goal of VinDr development team. VinDr is a comprehensive AI-based medical imaging solution developed by VinBigData, where Dung is working as the Medical Image Analysis team leader.
Put into pilot implementation in mid-2020, with the chest X-ray diagnostic module, VinDr is capable of detecting 22 other types of lesions such as effusion, pneumothorax, mass nodule, solidification, infiltrate, etc. The rate of accuracy is higher than 90%, making VinDr an useful assistant helping radiologists diagnose most of the lung-related diseases.
VinDr is currently being deployed in many hospitals and clinics in a number of provinces and cities nationwide. In addition, the development team has also completed 4 other modules including VinDr SpineXR, VinDr Mammo, VinDr ChestCT and VinDr Liver CT; thus supporting the screening and diagnosis of Vietnamese dangerous and common diseases.
“In fact, the models I used in the competition have been integrated into VinDr ChestXR module. For each dataset, I only have to optimize different parameters. In another words, the module helped me improve the competition model. At VinBigdata we develop a medical imaging solution with the desire to contribute to fighting against the pandemic, meanwhile supporting doctors effectively“, said engineer Ba Dung.
Before winning this position, Ba Dung and other members of the center had won many prestigious international awards in the field of medical imaging such as 1st prize in Abnormal Image Detection in Endoscopy Videos
(EndoCV), 2020, Top 3 in Pulmonary Embolism Detection Challenge organized by the Radiological Society of North America (RSNA), 2020, No.1 in CheXpert competition organized by Stanford University in 2019.
SIIM-FISABIO-RSNA Covid-19 Detection is a global competition held on the Kaggle from May 2021 to August 10, 2021, with total prize pool of $100,000.