Covid-MANet: Multi-task attention network for explainable diagnosis and severity assessment of COVID-19 from CXR images

Highlights：

• The Covid-MANet improves generalization and interpretability for COVID-19 classification by introducing segmentation-based cropping and classification by proposed MA-DenseNet201 model outperforming state-of-the-art networks.

• Hybrid loss function is used for lung and infection segmentation task whereas weighted loss function for performance improvement in classification problem.

• Investigates the class-wise sensitivity analysis at various confidence threshold levels. Based on prior awareness of various class level accuracies, a weighted average ensemble approach (WAE) outperforms state-of-the-art models for all the classes.

• Finally, a gradient-weighted class activation mapping (Grad-CAM) is used for explainable diagnosis to generate a localization map for each disease type, investigates model interpretation in addition to COVID-19 infection map.

摘要

•The Covid-MANet is a single generic multi-task framework for automated lung segmentation, COVID-19 diagnosis, infection region quantification and severity assessment of COVID-19 into more specific levels as mild, moderate, severe and critical.•The Covid-MANet improves generalization and interpretability for COVID-19 classification by introducing segmentation-based cropping and classification by proposed MA-DenseNet201 model outperforming state-of-the-art networks.•Hybrid loss function is used for lung and infection segmentation task whereas weighted loss function for performance improvement in classification problem.•Investigates the class-wise sensitivity analysis at various confidence threshold levels. Based on prior awareness of various class level accuracies, a weighted average ensemble approach (WAE) outperforms state-of-the-art models for all the classes.•Finally, a gradient-weighted class activation mapping (Grad-CAM) is used for explainable diagnosis to generate a localization map for each disease type, investigates model interpretation in addition to COVID-19 infection map.