Pulmonary COVID-19: Learning Spatiotemporal Features Combining CNN and LSTM Networks for Lung Ultrasound Video...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Pulmonology

Background:

• Deep learning is crucial for developing Computer-Aided Diagnosis (CAD) systems.
• Lung ultrasound (LUS) imaging is a valuable tool for diagnosing respiratory conditions.
• Accurate and rapid COVID-19 diagnosis remains a global health priority.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for classifying LUS videos to diagnose COVID-19.
• To leverage Convolutional Neural Networks (CNN) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal dependency learning.
• To optimize the hybrid CNN-LSTM model using the Optuna framework.

Main Methods:

• A hybrid CNN-LSTM model was designed, integrating Xception (pre-trained on ImageNet) for spatial feature extraction and LSTM for temporal analysis.
• Model hyperparameters were optimized using the Optuna framework.
• The model processed sequences of 20 frames from LUS videos captured by convex transducers.

Main Results:

• The optimized hybrid CNN-LSTM model achieved an average accuracy of 93% and a sensitivity of 97% for COVID-19 detection.
• The proposed model outperformed traditional spatial-only deep learning approaches.
• Transfer learning using ImageNet pre-trained models yielded results comparable to models pre-trained on LUS images.

Conclusions:

• The hybrid CNN-LSTM model is a highly effective tool for COVID-19 diagnosis using LUS videos.
• This approach demonstrates the potential of deep learning in medical image analysis for respiratory diseases.
• The findings support the use of LUS classification as a significant tool in managing COVID-19 and other lung conditions.