Computer Vision News - October 2022

22 Oral Presentation correspondingly if a part of the bonemoves. We transform the movement of the bone to the soft tissue. That is a major technical innovation.” Xi will be presenting the work today - his first oral presentation paper. He describes the group’s method in more detail: “First, we want to learn the spatial correspondence between bony and facial structures. We use PointNet++ networks to extract the structural features from the bony and facial point set. We learn the local spatial features from the facial and bony models and then compute their similarities. Each facial point has its corresponding bony points, and all the corresponding points have a weight to contribute to that specific facial point. There is a point-to- point correspondence matrix to transform the effect from the bone to the face.” They use MLP to encode the bony movement into a local bony movement feature. Compared to previous methods, which encode the bony movements into a global vector that cannot be decoded locally, they encode it locally and regarding efficiency . It would take some time to prepare the model, and surgeons with hectic schedules would not be able to wait in front of a computer for a model to be printed. “Scientifically, it was sound, but in reality, it was not clinically practical,” James says. “That is why we started thinking about the deep learning method, which is where Dr. Yan and his group come in.” Pingkun’s group has been working on deep learning techniques, including direct image processing, image analysis, and image construction, for several years. The two groups have come together to work on this problem and have joint authorship of the paper. Daeseung and Xi are co-first authors, and James and Pingkun are co- corresponding authors. It is a marriage of the clinical and the technica l. “We propose a deep learning-based approach,” Pingkun tells us. “FEM works because it can model the correspondence well between the bone and the facial tissue. Our method uses the attention mechanism to establish a correspondence, so that deep learning knows how the facial tissue moves BEST OF MICCAI