Due to the highly infectious and long incubation period of COVID-19, detecting COVID-19 efficiently and accurately is crucial since the epidemic outbreak. We proposed a new detection model based on U-Net++ and adopted dense blocks as the encoder. The model not only detects and classifies COVID-19 but also segment the lesion area precisely. We also designed a two-phase training strategy along with self-defined groups, especially the retrocardiac lesion to make model robust. We achieved 0.868 precision, 0.920 recall, and 0.893 F1-score on the COVID-19 open dataset. To contribute to this pandemic, we have set up a website with our model (https://medchex.tech/).