Efficient Semi-automatic Segmentation-Labeling of Any Volumetric Medical Image

Regions of interest are often labeled in volumetric medical images either for research purposes or for diagnosis and treatment planning.However, labeling such segments manually is time-consuming and requires medical expertise, which makes it expensive.We design a novel semi-automatic 3D workflow which allows efficient segmentationlabeling of volumetric images.To this end, for a given 3D image we first manually label a subset of its 2D slices using MedSAM (a foundational model for segmenting any 2D medical image) via bounding-box prompting.Subsequently, we interpolate user-provided prompts for the remaining slices to automatically generate labels for them.This way, users can process a complete volumetric image while working on only a subset of its slices.We evaluate our method on the diverse set of medical image datasets from the Medical Segmentation Decathlon challenge.Our approach significantly reduces the labeling effort, around 67%, while only marginally reducing the segmentation accuracy compared to applying MedSAM slice-by-slice.Breaking out of the time-consuming slice-by-slice workflow with only a minor reduction in accuracy is a significant step in streamlining the process of semi-automatic labeling.