Drosophila heart 3D segmentation using LSTM neural network in optical coherence microscopy (Conference Presentation)

We developed a custom convolutional neural network combined with long short-term memory , named FlyNet 2.0, to achieve accurate drosophila heart segmentation in the cross-sectional videos of drosophila heart acquired by optical coherence microscopy system. Our first version of FlyNet uses regular CNNs to only extract 2D spatial information from individual frames of drosophila heart OCM videos. LSTM utilizes both spatial and time information to further improve the drosophila heart segmentation performance. 500 thousands fly heart OCM images were employed to train and test with FlyNet 2.0. Drosophila heart OCM videos through multiple heartbeat cycles and different heartbeat dynamic recovery situations were tested with an intersection over union accuracy at 92%. With increased segmentation accuracy, dynamical cardiac parameters, like heart area, end diastolic diameter, and end systolic diameter, can be quantified more accurately for the study of drosophila cardiac disease.