Herbarium is considered to be a stiff problem among many competitions at this year’s CVPR conference. The team from Medical Imaging Department including 3 engineers Nguyen Thanh Nhan, Tran Quang Dat and Ngo Thanh Dat successfully built a model capable of classifying more than 32,000 different herbs with the data set of over a million photos within 2 months. As a result, the system ensured an accuracy of 84.5%, once again confirming the ability to master global-standard core technologies of VinBDI in particular and the next generation of Vietnamese scientists in general.
About the reasons for choosing herb classification – a problem that seems not related to medical imaging, Nguyen Thanh Nhan – the team leader shared: “Herb classification is an interesting problem. It is difficult to classify and find out the characteristics of each herb. With the large amount of data provided by organizer, this is an opportunity for us to take our algorithm to a world-scale competition and finally provide the best model. The achievement is also an evidence proving technological competence of the core algorithm that Medical Imaging Department is researching and developing.”
The algorithm consists of 5 convolutional neural networks (EfficientNet B3 / B4 / B5, Inception and ResNet) with many improvements to fit the fine-grained visual categorization. The experience gained from competition will be very helpful in expanding detailed diagnosis of lesions and pathologies. For example, calcification on mammograms can be divided into suspected microcalcifications, typical microcalcifications, mammary microcalcifications, and skin microcalcifications. The detailed diagnosis of lesions will provide more accurate information and better support for radiologists. More details on the model can be found at the link. The team is expected to provide open source and algorithm models to the community in the near future.
In addition to the latest award at this year’s CVPR Workshop, the Medical Imaging Department has repeatedly had high positions in 05 world-class Medical diagnostic imaging competitions in 2019 and 2020, including:
- Rank first in CheXpert competition organized by Stanford University (the largest competition for automated chest x-ray interpretation, which features uncertainty labels and radiologist-labeled reference standard evaluation sets).
- First prize in the ontest of abnormalities on endoscopic images of ISBI conference in 2020 (the world’s leading conference on medical imaging)
- Rank first (on the first phase) over 1400 teams in pneumothorax challenge organized by the Society for Imaging Informatics in Medicine (SIIM)
- Top 10 out of 1,345 organizations participated in the brain diagnostic imaging competition organized by RSNA (Radiological Society of North American – the largest medical imaging association in the world with 54,000 experts / doctors from 136 countries worldwide)
Based on these algorithms, the Medical Imaging Department is gradually finalizing and launching VinDr – A Comprehensive AI solution for medical imaging, helping doctors to make fast and precise decisions, supporting the detection and avoid missing any lesions. Focusing on ubiquitous pathological diseases in Vietnam, VinDr is developing 06 modules: radiological diagnosis in lung diseases; diagnosis of breast cancer on X-ray images; diagnosis of lung cancer on computed tomography (CT), diagnosis of liver cancer on CT, diagnosis of brain stroke on CT and diagnosis of brain tumors on magnetic resonance imaging.
The first version of that solution will focus on two modules: X-ray of lung diseases and X-ray of breast cancer. It is expected that the pilot implementation will begin in June 15, 2020 in 3 hospitals: 108 Military Central Hospital, Hanoi Medical University Hospital and Vinmec Times City International Hospital.
CVPR (Computer Vision and Pattern Recognition) is the premier annual computer vision event comprising the main conference and several co-located workshops and short courses. The conference is organized by IEEE Computer Society and CVF – The Computer Vision Foundation. It was first held in Washington DC in 1983. Now it is considered as one of the world’s leading conferences on Computer Vision and Pattern Recognition.
