Computer Vision News - August 2022

37 Differentiable Boundary Point... “ Before, the literature was more about trying to find some hacks to make this work within the current framework, ” Robin adds. “ I had this idea that if we get the boundary points, we will be able to get the diameter. It was a lot of work on blackboards, just trying to derive the equations to arrive at something we were happy with. Ultimately, that led towards convergence because if you have a strong mathematical model, you have more chance of converging. ” Throughout this project, Robin used a combination of MLflow for tracking experiments, PyTorch Lightning to ensure a standardized approach, and MONAI by NVIDIA , a toolbox for building computer visionmodelstargetedmoretowardmedical imaging. MONAI will be very familiar to our readers, and we must congratulate Prerna Dogra , Stephen Aylward , and the many scholars who have brought it to where it is today as such a valuable tool for scientists and engineers. Robin says it streamlined the process of creating the model and made the research phase of this work much faster . In this paper, Robin and the team try to predict a carotid artery segmentation based on its diameter annotations. The novelty of the work is a differentiable loss going directly from the output of the segmentation network to the boundary point coordinates and later in the pipeline to the diameter, which allows optimization downstream of a segmentation network based only on diameter annotation. “ Usually, in computer vision literature, you have a problem, and you create more and more models that are more and more complex using more and more data, ” Robin explains “ We wanted to use prior knowledge in our model to reduce this complexity and make sense of our methods. ” The model helps you to understand what you are optimizing based on the mathematical model integrated within the framework by adding this prior knowledge, which makes it more interpretable.