Alert button

Med-Query: Steerable Parsing of 9-DoF Medical Anatomies with Query Embedding

Dec 05, 2022
Heng Guo, Jianfeng Zhang, Ke Yan, Le Lu, Minfeng Xu

Figure 1 for Med-Query: Steerable Parsing of 9-DoF Medical Anatomies with Query Embedding
Figure 2 for Med-Query: Steerable Parsing of 9-DoF Medical Anatomies with Query Embedding
Figure 3 for Med-Query: Steerable Parsing of 9-DoF Medical Anatomies with Query Embedding
Figure 4 for Med-Query: Steerable Parsing of 9-DoF Medical Anatomies with Query Embedding

Share this with someone who'll enjoy it:

Automatic parsing of human anatomies at instance-level from 3D computed tomography (CT) scans is a prerequisite step for many clinical applications. The presence of pathologies, broken structures or limited field-of-view (FOV) all can make anatomy parsing algorithms vulnerable. In this work, we explore how to exploit and conduct the prosperous detection-then-segmentation paradigm in 3D medical data, and propose a steerable, robust, and efficient computing framework for detection, identification, and segmentation of anatomies in CT scans. Considering complicated shapes, sizes and orientations of anatomies, without lose of generality, we present the nine degrees-of-freedom (9-DoF) pose estimation solution in full 3D space using a novel single-stage, non-hierarchical forward representation. Our whole framework is executed in a steerable manner where any anatomy of interest can be directly retrieved to further boost the inference efficiency. We have validated the proposed method on three medical imaging parsing tasks of ribs, spine, and abdominal organs. For rib parsing, CT scans have been annotated at the rib instance-level for quantitative evaluation, similarly for spine vertebrae and abdominal organs. Extensive experiments on 9-DoF box detection and rib instance segmentation demonstrate the effectiveness of our framework (with the identification rate of 97.0% and the segmentation Dice score of 90.9%) in high efficiency, compared favorably against several strong baselines (e.g., CenterNet, FCOS, and nnU-Net). For spine identification and segmentation, our method achieves a new state-of-the-art result on the public CTSpine1K dataset. Last, we report highly competitive results in multi-organ segmentation at FLARE22 competition. Our annotations, code and models will be made publicly available at: https://github.com/alibaba-damo-academy/Med_Query.

* 13 pages, 12 figures  
View paper onarxiv icon

Share this with someone who'll enjoy it: