Fig. 2

MVnet pipeline. a The input cine images with an inherent clinical variability (size \(m \times n\), resolution, orientation and cropping) were fed to the proposed dual-stage residual neural network. b The first trained ResNet-50 produced coarse annotations, marked in circumferences representing acceptable accuracy, in every cine image in a fixed image size of 160 \(\times\) 160, which in turn, served to apply a c linear transformation to a standard spatial resolution of 0.75 mm, orientation and cropping around the mitral valve center for a size of 118 \(\times\) 162. d The second trained ResNet-50 used the transformed images to predict precise annotations, marked in circles representing higher accuracy, which were adjusted again to the original input image. These last two tasks could be done iteratively as indicated. e The output time-resolved coordinates were used to derive the mitral valve displacement and velocity curves