“An Attention LSTM U-Net Model for Drosophila Melanogaster Heart Tube Segmentation”
Thesis lab: Chao Zhou (WashU Biomedical Engineering)
Abstract: Machine learning is commonly used in biomedical image analysis as it allows automated image segmentation and identification that minimizes the need for tedious human involvement. When researching the cardiac function of the drosophila melanogaster, human experts were involved in manually identifying the heart position in optical coherence microscopy (OCM) images and analyzing its beating dynamic. As OCM often generates large volumes of images, automated image segmentation is necessary to efficiently quantify heart beating. Our most recent heart segmentation model, FlyNet 2.0+, is a fully convolutional LSTM U-Net model [1]. However, the performance of the model diminishes in the presence of artifacts, such as image reflection and heart movement, resulting in manual intervention for mask correction, which is time-consuming. Therefore, we developed the FlyNet 3.0 model with integrated attention gates in skip connections between each level of the LSTM U-Net model [2]. The attention model adaptively adjusted and automatically learned to focus on the target structure, the heart area. Compared to the previous model, Flynet 3.0 increases the prediction IOU accuracy from 0.86 to 0.89 for images with reflection artifacts and from 0.81 to 0.89 for those depicting heart movement. Furthermore, we have expanded the functionalities of OCM analyses through automated and dynamic heart wall thickness measurements, which we have validated on a drosophila model of cardiac hypertrophy.
For inquiries contact Aaron Beagle at abeagle@wustl.edu.