Intelligent Emboli Detection from Doppler Ultrasound Audio Recordings with Deep Learning

Ultrasonic Doppler examinations have been employed for evaluating the health status of divers post underwater activities, aiming to mitigate the risk of decompression sickness (DCS). However, traditional assessments have heavily relied on subjective judgements rendered by medical professionals. To enhance objectivity in evaluations, this study introduces a novel approach for automatic grading of Doppler bubble signals utilizing deep learning models and Mel Spectrogram analysis. For the purpose of algorithm training and validation, a synthetic dataset consisting of 15,000 10-second recordings of Spencer was enrolled. This dataset encompassed 1000 instances of Spencer categorized into five different grades, each grade comprising three variations (precordial full cardiac cycles, precordial partial cardiac cycles, and subclavian). Various signal preprocessing techniques, including empirical mode decomposition (EMD), high pass filters, were scrutinized. Furthermore, diverse neural network architectures, such as Res Net (ResNet50) and Swin Transformer v2-Small model, were assessed. The performance of the Swin Transformer v2-Small model proved noteworthy across various datasets, achieving an accuracy of 81.13 % for Doppler ultrasound (DU) signals' Mel spectrogram through filtering. Conversely, the ResNet50 model exhibited marginally lower overall performance, with the highest accuracy of 77.47% achieved on the original Mel Spectrogram. The findings suggest that employing the Swin-transformer v2 -s for classifying mel spectrum data subsequent to filtering is conducive to assessment based on stress disease risk levels.