来源：专知
本文约1000字，建议阅读5分钟
在这篇论文中，我提出了一种利用解剖学先验知识进行结构分割的方法，同时保持已知的拓扑结构。
Since the rise of deep learning, new medical image segmentation methods have rapidly been proposed and shown promising results, each reporting minor improvements over previous state-of-the-art methods. But in visual inspection, errors are often found, such as topological errors (e.g. holes or folds), which are not detected when using traditional evaluation metrics such as Dice. Furthermore, correct topology is often necessary in ensuring that segmentations are both anatomically and pathologically plausible and ultimately suitable for downstream image processing tasks. Therefore, it is necessary to focus on ensuring that the predicted segmentation is topologically correct, rather than just optimizing pixel-level accuracy. In this paper, I propose a method for structural segmentation that leverages anatomical prior knowledge while maintaining known topology.
The proposed model, Topology Encourages Deformable Segmentation Network (TEDS-Net), performs segmentation by deforming a previous shape with the same anatomical properties of interest, using a learned topology-preserving deformation field. However, I show here that such fields can only be topologically preserved in continuous domains, and when applied to discrete spaces, their properties start to break down. To overcome this effect, I introduce additional modifications in TEDS-Net to more strictly enforce topology preservation, a step often overlooked in previous work.
In this paper, TEDS-Net is applied to a series of natural and medical image segmentation tasks. I show how it works with several topology types, with several structures, and in 2D and 3D. Further, I show how to use TEDS-Net to segment the entire volume using minimally annotated training data. In these experiments, TEDS-Net topologically outperforms all SOTA baselines while maintaining competitive pixel-level accuracy.
Finally, TEDS-Net is integrated into an entire medical imaging pipeline to illustrate the importance of topologically correct segmentation in downstream tasks. TEDS-Net was used to segment the developing cortical plate in 3D intrauterine fetal brain ultrasound scans to enable the description of complex growth and development during pregnancy. As far as I know, this task has only been attempted before in magnetic resonance imaging (MRI), although ultrasound is the preferred modality in prenatal care. This may be due to large acoustic shadows blocking critical brain regions on ultrasound. Due to the anatomical limitations of TEDS-Net, it can anatomically guide cortical plate segmentation in shaded regions, yielding a complete segmentation that enables accurate downstream analysis.
This is a rough translation that covers the core meaning of the original text. For professional papers or publication purposes, further proofreading and refinement of this translation is recommended to ensure accuracy and fluency.
