In our advanced machine learning course, we worked in a group of three students to develop a project aimed at performing image segmentation on echography videos. Our goal was to draw areas where parts of the mitral valve are. As only a few pictures were manually drawn by a medical doctor, the remaining picture has to be automatically drawn. This automatisation makes the process faster, more affordable, and less laborious.
Segmenting echographies is a challenging task due to the low image quality of this procedure and the deforming nature of organs. We decided to tackle this problem with deep learning and trained an artificial intelligence to segment the images. To ensure high-quality training, we customized the dashboard according to our requirements.
We also trained another artificial intelligence to localize the organ of interest with a rectangular bounding box. Combining localization and segmentation enhanced the output's quality. For better generalization, we performed live data augmentation with deforming the images slightly during training.