Sammendrag
The timing of cardiac events is essential for the analysis of certain components of myocardial function [36]. Finding an algorithm that detects these timings has therefore been the subject of several studies[36]. In this project, deep neural networks were used to detect the valvular event times from echocardiography sequences. Three classes of neural network algorithms were tested: fully convolutional architectures [48], VGG [47] inspired architectures and recurrent neural networks. Temporal information was also incorporated by feeding the network the relative time passed since the last QRS peak. It was found that incorporating temporal information was necessary for detecting the valvular event times with an acceptable accuracy. The model providing the highest performance metrics was a VGG inspired architecture with both an RNN head and the relative time since the QRS peak. A version of this model that did not include the relative times was visualised using both guided backpropagation [48] and image occlusion [54], which demonstrated that the position and movement of the valves were important for correctly predicting valvular events. The best model achieved a 93% test accuracy and correctly detected all the valvular events in 7 out of 11 test series with a mean error of 1.03 frames. This is not satisfactory for clinical use. However, it does indicate that deep neural networks applied to echocardiography are a promising approach for automatic valvular event time detection