The development of the biomedical industry has always been highly regarded by the Executive Yuan. For example, precision medicine has been featured as one of the key industries in the Biomedical Industry Innovation Program launched in 2017. Besides, the precision health industry, which emphasizes on smart healthcare, is also categorized as one of the current-planned six core strategic industries. Nowadays, artificial intelligence (AI) has been widely applied in the fields of medical and healthcare thanks to the advancement of computers and the rapid development of AI technologies such as deep learning. AI is the most important tool for big data mining which enables not only rapid discovery for insights and novel solutions from big data of healthcare, but also analysis to comprehensive information in real time for disease prognosis to support medical decision-making. Acknowledging that data is an indispensable element for the development of medical AI, the Ministry of Health and Welfare, the Ministry of Science and Technology, and the Ministry of Economic Affairs will jointly cooperate to initiate a four-year “Sustainable Big Data Platform for Precision Health” project in 2021 with the goals of establishing a big data platform for precision health which serves as the foundation of sustainability of the data, and introducing innovative business model by using complex big data to achieve value-added clinical translation research.
Due to the fact that medical imaging has long been an indispensable method in medical diagnostics, it is the earliest emerging, and has always been the most popular development field of smart healthcare. With the growing geriatric population and the subsequent increase in the prevalence of cancer and chronic diseases, demands for medical imaging increase dramatically and hence the workloads for physicians. Driven by the needs for AI-aided image diagnostics, the Ministry of Science and Technology sponsored three R&D teams in National Taiwan University, Taipei Veterans General Hospital, and Taipei Medical University to develop several distinct diagnosis tools for medical imaging focusing on the major heart-, brain-, and lung-related diseases which could be commonly diagnosed among Taiwanese.
Taipei Veterans General Hospital (TVGH) and Taiwan AI Labs have been jointly developing AI application in medical service since 2017. DeepMets®, an example of AI-assisted Imaging diagnostic model, is currently deployed in neuroimaging and chest medicine outpatient clinics at TVGH. The model training started at the early stages by adopting therapeutic-level annotated brain magnetic resonance (MR) images of more than 1,000 patients, who have cancers that metastasize to the brain. To complete the training and optimization of the AI model, DeepMets® leveraged “Harmonia,” a Taiwan-developed and open-sourced federated learning (FL) framework, and incorporated more than 3,000 patients’ brain MR images from National Health Insurance Administration’s imaging database, containing data from 23 medical centers. Currently, this application can accurately identify lesions of brain metastasis and calculate lesion quantity, maximum diameter and volume based on the MR images produced by major international brands such as GE, Siemens, Philips and so on. DeepMets® has assisted in diagnosis of more than 1,500 cases since 2019, and reduced the time physicians required to review the medical images from 10 minutes to only 30 seconds and achieved a F1-score of 96. Patient’s medical service process is feasibly shortened from 2 weeks to a half day by the setting. The application has facilitated decision-making for treatment and alleviates patient’s anxiety along the process.
The team of Taipei Medical University focused on the application for lung-related diseases. By adopting a huge amount of deep-annotated CT images, semantic features, and pathological slides, a novel AI system was developed with dual main engines specifically for (1) lung nodule detection, classification (malignant/benign), automatic radiology report generation through CT images and for (2) cancer areas detection and cell analysis calculation through digital automatic pathology slides. The system is able to quickly screen low-dose lung computerized tomographic (CT) images, detect and analyze spots and characteristics of lung nodule, and within 10 seconds automatically generate LungRads evaluation reports meeting international standards. The system helps a radiologist to largely reduce the review time from 20~30 minutes to less than 5 minutes for a case and within a precision rate of 95% or above. It can greatly improve the efficiency and error rate of radiologists from large and medium-sized medical institutions. In addition, the whole slide pathological detection of the system can accelerate the pathological diagnosis of lung cancer and assist in quantitative analysis of subtype lung adenocarcinoma.
The R&D team from National Taiwan University/National Taiwan University Hospital (NTUH) focused on Cardiovascular applications. The AI model developed jointly with NVIDIA by adopting high-quality annotation of pericardium and ascending/descending aorta contour in chest CT images, verified by cardiac/medical imaging physicians from 8 major medical centers, and the self-developed neural network architecture, with federated learning of images from multiple centers and different CT machines, making it the one and only AI model (HeaortaNet®) in the world that can automatically categorize and calculate the quantification of chest calcification/fat. The accuracy of heart segmentation can reach 94.2%, and it only took 0.4 seconds to analyze one case. At present, this Model has been certified by NVIDIA and placed on the NGC (NVIDIA GPU Cloud) for global AI research use. It has also been implemented and exercised in the NTUH’s Department of Medical Imaging and the Big Database Platform of NHI in >5,000 cases to assist in reporting and building risk models for prognosis of cardiovascular disease for Taiwanese domestic.
The Ministry of Science and Technology promotes the combination of medical and AI by funding the development of AI systems to assist doctors in diagnosing diseases, which serves as demonstration cases for value-added medical imaging in precision medicine. An image annotation standard has been established to guide the preparation of data for building machine learning models. Following the standard, a cross-institution team of medical professionals built a high-quality annotated medical image database. A data-sharing mechanism that complies with the privacy-protection regulations has also been proposed for data reuse. Besides, a data-sharing platform along with the cloud data utilization mechanism has been built at the National Center for High-performance Computing of NARLABs. The aforementioned outcomes of the project, including high-quality annotated medical image database and the data-sharing mechanism, serve as a strong foundation for the development of new data-driven smart healthcare technologies and medical AI technologies in Taiwan.
Dr. Tzy-Mei Lin
Department of Foresight and Innovation Policies, Ministry of Science and Technology