Automatic Scoring of COVID-19 LUS Videos Using Cross-correlation-Based Features Aggregated from Frame-Level Confidence Levels Obtained by a Pre-trained Deep Neural Network

The emergence of COVID-19 has encouraged researchers to seek a method to detect and monitor patients infected with SARS-CoV 2. The use of lung ultrasound (LUS) in this setting is rapidly spreading because of its portability, cost-effectiveness, real-time imaging, and safety. LUS has demonstrated the potential to be widely used to assess the condition of the lungs in COVID-19 patients. Given frame-level labels provided by a pre-trained deep neural network (DNN), our goal is to identify an aggregation strategy that allows to move from framelevel to video-level, which is the output required by physicians for clinical evaluation. To achieve this goal, we propose a novel aggregation method based on the cross-correlation coefficients. The logic behind this idea is that, based on the similarity between the score's variables (at frame level), the cross-correlation should be informative as to how to discriminate at video level. We applied our approach to the LUS data from a multi-center study comprising of 283, 231, and 448 LUS videos from Lodi General, Gemelli, and San Matteo Hospital, respectively. Results show that the video-level agreement with clinical experts is obtained in 87.6% of the cases, which represents a promising video-level accuracy.

Automated summary

The emergence of COVID-19 has prompted researchers to use lung ultrasound to detect and monitor patients infected with SARS-CoV 2. A novel aggregation method based on cross-correlation coefficients allows for accurate assessment of the condition of COVID-19 patients' lungs at the video level, providing valuable information for clinical evaluation.