Objectives: Echocardiography plays a central role in the identification and treatment of heart disease and image interpretation requires extensive experience and time. Deep learning techniques can identify complex patterns from large labeled datasets and have revolutionized image classification, object detection and segmentation. To date, state-of-the-art automated image interpretation has not been widely applied to echocardiogram interpretation. The first cognitive step in image interpretation is imaging view identification. The aim of this study is to create an automated processing pipeline for determining the echocardiographic imaging view and evaluate the accuracy of various deep learning algorithms.
Methods: Individual parasternal long-axis (PLAX) and non-PLAX (other) DICOM images were obtained and de-identified for testing. Images were individually sorted based on imaging view (PLAX vs other) by a board certified echocardiographer. Deep convolutional neural networks were trained to sort images in a similar fashion. The algorithm was trained on an 80% sample of images and accuracy was tested on the remaining 20%. The accuracy of each network architecture was compared to view identification by a blinded echcoardiogapher. Three separate network architectures (LeNet, VGG-16, and DenseNet) were implemented using the TensorFlow API in python. Input images to each network were resized to 224x224 pixels to balance resolution, memory, and compute.
Results: 42,459 individual parasternal long-axis (PLAX) and 301,557 non-PLAX (other) DICOM images were used for analysis. The accuracy of the LeNet (11.3 million parameters), VGG-16 (40.4 million parameters), and DenseNet (7 million parameters) networks on the validation set of echocardiogram images was 85.0%, 97.50%, and 99.94% respectively for image identification. The compute time for forward pass inference is comparable for each network taking only milliseconds.
Conclusion: Vendor independent deep learning networks can rapidly and accurately identify features on standard echocardiogram images. DenseNet network architecture matched human level performance. Deep learning has the potential for rapid, automated image interpretation and can improve the accuracy and efficiency of echocardiogram interpretation.