eduzhai > Applied Sciences > Engineering >

Learning Diagnosis of COVID-19 from a Single Radiological Image

  • Save

... pages left unread,continue reading

Document pages: 17 pages

Abstract: Radiological image is currently adopted as the visual evidence for COVID-19diagnosis in clinical. Using deep models to realize automated infectionmeasurement and COVID-19 diagnosis is important for faster examination based onradiological imaging. Unfortunately, collecting large training datasystematically in the early stage is difficult. To address this problem, weexplore the feasibility of learning deep models for COVID-19 diagnosis from asingle radiological image by resorting to synthesizing diverse radiologicalimages. Specifically, we propose a novel conditional generative model, calledCoSinGAN, which can be learned from a single radiological image with a givencondition, i.e., the annotations of the lung and COVID-19 infection. OurCoSinGAN is able to capture the conditional distribution of visual finds ofCOVID-19 infection, and further synthesize diverse and high-resolutionradiological images that match the input conditions precisely. Both deepclassification and segmentation networks trained on synthesized samples fromCoSinGAN achieve notable detection accuracy of COVID-19 infection. Such resultsare significantly better than the counterparts trained on the same extremelysmall number of real samples (1 or 2 real samples) by using strong dataaugmentation, and approximate to the counterparts trained on large dataset(2846 real images). It confirms our method can significantly reduce theperformance gap between deep models trained on extremely small dataset and onlarge dataset, and thus has the potential to realize learning COVID-19diagnosis from few radiological images in the early stage of COVID-19 pandemic.Our codes are made publicly available atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×