Study on Performance Improvement of AI for Computer Aided COVID-19 Diagnosis by Adding Large Number of Negative Data

Abstract

We report on the construction of a machine to discriminate positive/negative for COVID-19 based on chest X-ray CT images. All slices of the input CT images are input to the same encoder and are identified as positive/negative for COVID-19 based on their maximum values. In this paper, we report the results of evaluating the impact of（1）different encoders,（2）the presence or absence of segmentation of the lung field region, and（3）the presence or absence of a large amount of additional negative data on the discrimination performance. In particular, although the negative data are collected in overwhelmingly greater numbers than the positive data, simply adding the negative data to the training data causes a significant imbalance in the number of positive and negative data, which hampers the learning of the discriminator. Therefore, we introduce a method to correct the imbalance in the number of data based on the margin and utilize the negative data to improve the discriminator's performance. In this paper, we explain the outline of how to deal with the imbalance in the number of discriminant data and report the experimental results.