Cardiac Phase Recognition in 2D Echocardiography through Integrated Transfer Learning within Autoencoder and Recurrent Neural Networks

This study addresses the challenge of detecting end-systolic (ES) and end-diastolic (ED) frames in echocardiography sequences, which is crucial for assessing cardiac function. We propose an integrated approach that combines a ResNet-based autoencoder for image feature extraction with recurrent neural networks (RNNs) for frame labeling. The autoencoder, optimized with Sobel and MSE loss functions, processes echocardiography sequences to reconstruct despeckled images. Based on transfer learning, the features from the autoencoder are then fed into an RNN model with Long Short-Term Memory (LSTM) and dense layers for frame label prediction. By using datasets including EchoNet, EchoNetpediatric, CAMUS, and PACS data, our model achieved an average absolute frame difference (aaFD) of 2.61 and 2.12 frames for ED and ES in apical four- chamber (A4C) view, and aaFD of 3.12 and 2.54 frames in apical two- chamber (A2C) view, respectively. This method demonstrates a promising step towards automated cardiac phase recognition in 2D echocardiography sequences.