According to an abstract posted on arXiv, current state-of-the-art segmentation models still produce anatomically implausible shapes on the structures that are hardest to segment, including broken vessels and fragmented glands. The authors identify this as a persistent problem with existing approaches.
The proposed solution, called DART, is a learned shape representation that scales across anatomical classes and imaging modalities. The abstract indicates that the representation captures what is necessary to enforce plausible shapes during segmentation.
The paper is available at arXiv:2610.27605. The abstract text is short, and no further details on the method's implementation or experimental results are provided in the source. Thus, only the problem statement and the general design goal of DART are reported here.