Measurements of dimensions, e.g., of left ventricle in two-dimensional echocardiography (2DE), can be highly significant markers of several cardiovascular diseases and such measurements are often used in clinical care.
The location of positions needed for measurements is however challenging in ultrasound images due to fuzzy boundaries and varying reflection patterns between frames. This also leads to a large variability between observers, where effective automation may be an approach to reduce this variability. However, due to a need for expert annotations representing a variation of diagnoses and due to observer variability, training data are scarce and can be noisy. The aim is therefore to develop methods and approaches enabling training of deep networks from these limited annotations.
For the problem of automatic measurements of dimensions of the left ventricle in two-dimensional echocardiography, Vingmed has a substantial collection of image data available. Some of these are labelled by experts and can be used as a starting point for investigating approaches for learning from limited data. We will start from this and analyze both the data and the amount and quality of labels to decide on potential methodological directions.
At VI seminar #3 Andrew Gilbert presented his research on deep learning and ultrasound imaging. The presentation is available below and concerns the issue of small datasets in medical imaging. They propose to generate synthetic labeled data for ultrasound imaging using a generative adversarial network (GAN).