Visual Sound Source Separation With Partial Supervision Learning

Since computed tomography (CT) provides the most sensitive radiological technique for diagnosing COVID-19, CT has been used as an efficient and necessary aided diagnosis. However, the size and number of publicly available COVID-19 imaging datasets are limited and have problems such as low data volume, easy overfitting for training, and significant differences in the characteristics of lesions at different scales. Our work presents an image segmentation network, Pyramid-and-GAN-UNet (PGUNet), to support the segmentation of COVID-19 lesions by combining feature pyramid and generative adversarial network (GAN). Using GAN, the segmentation network can learn more abundant high-level features and increase the generalization ability. The module of the feature pyramid is used to solve the differences between image features at different levels. Compared with the current mainstream method, our experimental results show that the proposed network achieved more competitive performances on the CT slice datasets of the COVID-19 CT Segmentation dataset and CC-CCII dataset.