An integrated autoencoder-based hybrid CNN-LSTM model for COVID-19 severity prediction from lung ultrasound

The COVID-19 pandemic has become one of the biggest threats to the global healthcare system, creating an unprecedented condition worldwide. The necessity of rapid diagnosis calls for alternative methods to predict the condition of the patient, for which disease severity estimation on the basis of Lung Ultrasound (LUS) can be a safe, radiation-free, flexible, and favorable option. In this paper, a frame-based 4-score disease severity prediction architecture is proposed with the integration of deep convolutional and recurrent neural networks to consider both spatial and temporal features of the LUS frames. The proposed convolutional neural network (CNN) architecture implements an autoencoder network and separable convolutional branches fused with a modified DenseNet-201 network to build a vigorous, noise-free classification model. A five-fold cross-validation scheme is performed to affirm the efficacy of the proposed network. In-depth result analysis shows a promising improvement in the classification performance by introducing the Long Short-Term Memory (LSTM) layers after the proposed CNN architecture by an average of 7 - 12%, which is approximately 17% more than the traditional DenseNet architecture alone. From an extensive analysis, it is found that the proposed end-to-end scheme is very effective in detecting COVID-19 severity scores from LUS images.

Automated summary

This paper proposes a disease severity prediction architecture based on Lung Ultrasound by integrating deep convolutional and recurrent neural networks to consider spatial and temporal features of LUS frames, significantly improving the classification performance of COVID-19 severity scores.