Computer Vision News - May 2020

3 Summary Deep Learning on Echoc dio 11 computer vision algorithms to aid the assessment phase in Echocardiography. In particular, as with many other medical imaging modalities, the recent deep learning revolution in computer vision holds significant potential for Echocardiography. At the most basic level deep learning can be used to help organize and classify the huge amount of data produced in studies. Deep convolutional neural networks (CNNs) can be trained to automatically classify clips according to view, and even to provide a quality score for each clip. At a higher level, CNNs can be developed to segment the various cardiac structures, and to produce corresponding measurements. For example, the area of the left ventricle output tract can be measured, which is needed to calculate the cardiac output, or the area of the ventricle cross-section can be monitored through the cardiac cycle to provide the fractional area change between diastole and systole. At an even higher level CNNs can be trained to identify and classify specific cardiac pathologies to support the diagnostic procedure. Our experience at RSIP Vision shows that a critical element of developing a successful deep learning model for Echocardiography is high quality annotation of the training data. To achieve this the annotators should have some medical background, and should be trained, supervised and audited by a highly experienced echo specialist. Even then, the annotation and training processes may require a number of iterations until satisfactory results are achieved. Regarding the actual CNN models themselves, we have found that 3D models which take as input an entire clip, or at least a subset of frames from a clip, are much more effective. This is true even in cases where the analysis could in theory be done using a 2D model taking a single frame as input. To summarize, Echocardiography is an indispensable tool for cardiovascular disease diagnosis and management, and deep learning holds significant promise for making this tool even more effective and efficient. In order to realize this potential and develop high quality and validated deep learning models which can be used in real-world clinical situations, a high level of expertise and careful attention to detail is required. RSIP Vision is the ideal partner to take along in all your deep learning projects.