ADDG: An Adaptive Domain Generalization Framework for Cross-Plane MRI Segmentation

Zibo Ma, Bo Zhang, Zheng Zhang, Wu Liu, Wufan Wang, Hui Gao, Wendong Wang

Abstract: Multi-planar and multi-slice magnetic resonance imaging (MRI) can provide more comprehensive 3D structural information for disease diagnosis. However, compared to multi-source MRI, multi-planar MRI uses almost the same scanning parameters but scans different internal structures. This atypical domain difference may lead to poor performance of traditional domain generalization methods in handling multi-planar MRI, especially when MRI from different planes also comes from different sources. In this paper, we introduce ADDG, an Adaptive Domain Generalization Framework tailored for accurate cross-plane MRI segmentation. ADDG significantly mitigates the impact of information loss caused by slice spacing by incorporating 3D shape constraints of the segmentation target, and better clarifies the feature differences between different planes of data through adaptive data partitioning strategy. Specifically, we propose a mesh deformation-based organ segmentation network to simultaneously delineate the 2D boundary and 3D mask of the prostate, as well as to guide more accurate mesh deformation. We also develop an organ-specific mesh template and employ Loop subdivision for unpooling new vertices to a triangular mesh to guide the mesh deformation task, resulting in smoother organ shapes. Furthermore, we design a flexible meta-learning paradigm that adaptively partitions data domains based on invariant learning, which can learn domain invariant features from multi-source training sets to further enhance the generalization ability of the model. Experimental results show that our approach outperforms several medical image segmentation, single-planar-based 3D shape reconstruction, and domain generalization methods.

Primary Subject Area: [Content] Media Interpretation

Secondary Subject Area: [Experience] Multimedia Applications

Relevance To Conference: Domain Generalization is an essential problem for utilizing deep learning techniques in several applications such as multimedia data processing. Our work investigates a widespread but as yet unattended problem of domain generalization, filling the gap in domain generalization to some extent. This cross-view domain generalization problem is also prevalent in multimedia information processing, and thus research addressing this problem has the potential for future application to cross-view domain generalization problems in multimedia data processing.

Submission Number: 1309