An Analysis of Cavitation in Sonothrombolysis through Convolutional Neural Networks

Myocardial infarction is one of the main causes of morbidity and mortality worldwide. Among the possible treatments for blood flow obstruction, an emerging technique is named sonothrombolysis. To reach satisfactory results, the event allied to the technique (cavitation of microbubbles) needs to be controlled, so harm to the patient can be avoided. In view of that, this study aimed to detect and classify the phenomenon during sonothrombolysis therapy through the use of artificial intelligence. The signals were generated considering an 8 x 8 transducers' matrix, and an automatic and uncomplicated classifier method was proposed, based on the Continuous Wavelet Transform tool and Convolutional Neural Network (CNN) approach. The method made use of a pre-trained CNN, called AlexNet, operating a database of 2,800 waves for training, testing, and validation. The evaluation of the statistics included both the detection using broad and narrow bands, the noise level applied, and the database size. For the case of narrowband receivers, the results of the study indicated an accuracy of around 95.7%. The result demonstrates that the use of artificial intelligence could be an approach to explore the detection of cavitation for therapies applying ultrasound.