Deep Learning-Based Venous Gas Emboli Grade Classification in Doppler Ultrasound Audio Recordings

Abstract

Objective: Doppler ultrasound (DU) is used to detect venous gas emboli (VGE) post dive as a marker of decompression stress for diving physiology research as well as new decompression procedure validation to minimize decompression sickness risk. In this article, we propose the first deep learning model for VGE grading in DU audio recordings.

Methods: A database of real-world data was assembled and labeled for the purpose of developing the algorithm, totaling 274 recordings comprising both subclavian and precordial measurements. Synthetic data was also generated by acquiring baseline DU signals from human volunteers and superimposing laboratory-acquired DU signals of bubbles flowing in a tissue mimicking material. A novel squeeze-and-excitation deep learning model was designed to effectively classify recordings on the 5-class Spencer scoring system used by trained human raters.

Results: On the real-data test set, we show that synthetic data pretraining achieves average ordinal accuracy of 84.9% for precordial and 90.4% for subclavian DU which is a 24.6% and 26.2% increase over training with real-data and time-series augmentation only. The weighted kappa coefficients of agreement between the model and human ground truth were 0.74 and 0.69 for precordial and subclavian respectively, indicating substantial agreement similar to human inter-rater agreement for this type of data.

Conclusion: The present work demonstrates the first application of deep-learning for DU VGE grading using a combination of synthetic and real-world data.

Significance: The proposed method can contribute to accelerating DU analysis for decompression research.