UniPro
A unified multimodal medical image segmentation model that supports four modes within a single framework: semantic segmentation, contextual segmentation, interactive segmentation, and propagation-based segmentation. It handles both 2D images and 3D volumetric data, and improves volumetric propagation reliability through bidirectional and 3D supervision.
Timeline (last 90 days)
UniPro is a unified medical image segmentation framework that, within a slice-based framework, simultaneously supports four modes: semantic segmentation, contextual segmentation, interactive segmentation, and propagation-based segmentation.
The core insight of UniPro is that volumetric data propagation and contextual segmentation fundamentally share the same reference-conditioned prediction mechanism, differing only in the source of the reference image-mask pairs.
In volumetric data propagation, already-segmented adjacent slices play the role of reference examples, while the slice to be predicted plays the role of the query image.
UniPro's four segmentation modes share the same network architecture, differing only in the mode-specific conditional inputs fed to the model.
UniPro's semantic segmentation mode makes predictions using category priors as conditional inputs.
UniPro's interactive segmentation mode accepts user clicks as conditional inputs.
In addition to the conventional slice-by-slice loss, UniPro additionally introduces bidirectional supervision and 3D supervision to improve propagation reliability.
UniPro achieves strong performance across all segmentation settings in experiments covering multiple imaging modalities and different anatomical structures.
UniPro supports efficient complete 3D segmentation starting from sparse 2D initialization, reducing the workload of slice-by-slice correction for physicians.
UniPro is proposed as an arXiv preprint.
All Facts (10)
UniPro is a unified medical image segmentation framework that, within a slice-based framework, simultaneously supports four modes: semantic segmentation, contextual segmentation, interactive segmentation, and propagation-based segmentation.
50%UnverifiedThe core insight of UniPro is that volumetric data propagation and contextual segmentation fundamentally share the same reference-conditioned prediction mechanism, differing only in the source of the reference image-mask pairs.
50%UnverifiedIn volumetric data propagation, already-segmented adjacent slices play the role of reference examples, while the slice to be predicted plays the role of the query image.
50%UnverifiedUniPro's four segmentation modes share the same network architecture, differing only in the mode-specific conditional inputs fed to the model.
50%UnverifiedUniPro's semantic segmentation mode makes predictions using category priors as conditional inputs.
50%UnverifiedUniPro's interactive segmentation mode accepts user clicks as conditional inputs.
50%UnverifiedIn addition to the conventional slice-by-slice loss, UniPro additionally introduces bidirectional supervision and 3D supervision to improve propagation reliability.
50%UnverifiedUniPro achieves strong performance across all segmentation settings in experiments covering multiple imaging modalities and different anatomical structures.
50%UnverifiedUniPro supports efficient complete 3D segmentation starting from sparse 2D initialization, reducing the workload of slice-by-slice correction for physicians.
50%UnverifiedUniPro is proposed as an arXiv preprint.
50%