| Résumé | Cardiac murmurs are the first signs of pathological changes in heart valves. Their subtle presence poses great challenges for detection through auscultation or phonocardiograms (PCGs); therefore computer-aided detection (CAD) of heart murmurs has medical significance in assisting health care professionals. Traditional CAD approaches, relying on engineered features, are prone to changes in environmental noise and data collection methods. Deep Convolutional Neural Networks (CNN) have shown robustness in advancing the performance of computer vision tasks through automatic feature learning from large amount of data. Meanwhile, deep Recurrent Neural networks (RNN) have demonstrated state-of-the-art performance on processing sequence data in areas such as speech recognition and natural language processing. With a limited set of labelled PCG recordings, this work first transforms PCGs to topology-preserving spectral images, conducts data augmentation and successfully tunes latest deep CNN architectures with transfer learning. Experimental results indicate that the fine-tuned deep CNNs are effective in classifying cardiac murmurs from PCG recordings without segmentation. Moreover, the deep CNNs gained a further performance boost from learning the dynamics in temporal sequences after being plugged into an RNN model. In summary, the proposed deep recurrent-convolutional neural network approach captured spectral, temporal and spatial features. It achieved an overall accuracy of 94.01% for automatic cardiac murmur detections in phonocardiograms. |
|---|
