MICCAI 2019 Tuesday

MICCAI 2019 DAILY 11 Mathilde Bateson cancer is more difficult because you have less constraints. Cancer can have a lot of shapes. It can be a lot harder to input knowledge, whereas, the heart and maybe even the lungs have more definite shape. How is it different to analyze the vertebrae compared to other parts of the body? In my paper, I worked on domain adaptations. The purpose is to make models robust across different imaging modalities. For vertebrae, when you do an MRI of your back, it’s very common to have four modalities. You have four different images that have very different aspects. The reason for that is that clinicians look at different things. They can see different things from the four images. What is very important is to learn a model that works on every modality, that can extract the most information from every modality. For example, in domain adaptation, you want to have a model that works very well on one modality but also can transfer very well on another modality. This is the purpose of our work. How is it different to analyze the vertebrae compared to other parts of the body? In my paper, I worked on domain adaptations. The purpose is to make models robust across different imaging modalities. For vertebrae, when you do an MRI of your back, it’s very common to have four modalities. You have four different images that have very different aspects. The reason for that is that clinicians look at different things. They can see different things from the four images. What is very important is to learn a model that works on every modality, that can extract the most information from every modality. For example, in domain adaptation, you want to have a model that works very well on one modality but also can transfer very well on another modality. This is the purpose of our work. Tell us more about the poster you are presenting. The title is Constrained Domain Adaptation for Segmentation. We try to do what is called unsupervised adaptation. This means that you have two modalities. You have the source modality, which is fully supervised. You have all the annotations on the source image. On the target image, you have no annotation at all. This is why it’s unsupervised. This is a very hard task. Usually, the models drop in quality very harshly. What we try to do is integrate prior knowledge, something that is very easily accessible. We choose the size of the vertebrae. We chose the size of the structures as the prior knowledge available in the clinic and try to integrate it into the deep learning framework. We did this. We motivate