C2FNAS: Coarse-to-Fine Neural Architecture Search for 3D Medical Image Segmentation

Segmentation is an important research field in computer vision. It requires recognizing and segmenting the objects at the pixel level. In the past decade, many deep neural networks have been proposed, which have been central to the development in this area. These frameworks have demonstrated human-level or beyond performance on many challenging benchmarks, and have been widely used in many real-life applications, including surveillance, autonomous driving, and medical image analysis. However, it is non-trivial to design neural architectures with both efficiency and effectiveness, especially when they need to be tailored to the target tasks and datasets. In this dissertation, I will present our research works in this area from the following aspects. (i) To enable automatic neural architecture design on the costly 3D medical image segmentation, we propose an efficient and effective neural architecture search algorithm that tackles the problem in a coarse-to-fine manner. (ii) To further take advantage of the neural architecture search, we propose to search for a channel-level replacement for 3D networks, which leads to strong alternatives to 3D networks. (iii) To perform segmentation with great detail, we design a coarse-to-fine segmentation framework for matting-level segmentation; (iv) To provide stronger features for segmentation, we propose a stronger transformer-based backbone that can work on dense tasks. (v) To better resolve the panoptic segmentation