Early detection of infection Data consolidation/analysis for the development of a diagnostic-aid AI model using chest CT imaging

Research and Development by

Naoki Okada, Osaka General Medical Center / fcuro Co., Ltd.

Corresponding Research Area

BioMedical Counter-measures: Analysis of CT scans for early detection and avoidance of severe illness, monitoring of patients with mild illness, prediction of risk of severe illness, understanding of the effects of viral mutation, etc.

Minimizing physician man-hours and contact between coronavirus disease 2019 (COVID-19) patients and general patients/medical staff by early sorting of COVID-19 patients through artificial intelligence (AI)-assisted computed tomography (CT) diagnosis

[Previous initiatives]

Data structure survey of the accumulated datasets

Construction of a recognition model using open data
Normalization (refinement to absorb differences in characteristics between datasets)
Pre-processing (automatic extraction of lung fields to improve accuracy for the recognition of the judgment algorithm)
Construction of COVID-19 identification algorithm

Results of accuracy test for the recognition model with cases in Japan

A total of 1521 COVID-19, 1456 common, and 1240 normal cases were accumulated
CC-CC2 (data from all hospitals in China), severe acute respiratory syndrome coronavirus 2 CT-scan dataset (data from hospitals in Sao Paulo, Brazil), etc.

Normalization is required to create a common dataset because each dataset has different image characteristics

The open data-based identification algorithm showed: 87% accuracy for open data
Approximately 50% accuracy for cases in Japan

  • AI diagnosis systems are being developed with CT images from China, Brazil, etc.; however, because using these systems as is for cases in Japan delivers insufficient accuracy, examinations to improve their accuracy are ongoing.
  • An optimal AI model algorithm based on the analysis of data accumulation reflecting the real situation is essential for clinical settings in Japan

[Future plans]

Scheduled to be completed by the end of January 2021

  • Multicenter data accumulation of COVID-19 cases in Japan for algorithm testing with a larger accumulation of domestic data (approximately 246 institutions across Japan have been requested to provide data, as recommended by a Japanese Association for Acute Medicine study)
  • The AI diagnosis model is tested with domestic multicenter data, and improvement specifications are formulated to allow for practical use in medical settings in Japan

After February 2021, development will continue, and the algorithm will be deployed at clinical institutions in Japan for demonstration

Courtesy of Osaka General Medical Center, fcuro inc.