MID
A distribution-level evaluation metric for medical image segmentation that assesses segmentation quality by comparing two joint distributions, the reference set and the prediction set, without requiring sample-level pairing. It captures both mask realism and image-mask consistency.
Timeline (last 90 days)
Researchers at Caravel Lab proposed MID (Mask-Image Distributional Divergence), a distribution-level evaluation metric for medical image segmentation.
The official implementation of MID has been open-sourced at github.com/erc-caravel/pytorch-mid.
The starting judgment of MID is: segmentation quality is a property of the joint image-mask distribution.
MID directly compares two joint distributions, the reference set (real image-mask pairs) and the prediction set (image-predicted mask pairs), requires no sample-level correspondence at all, and simultaneously captures two dimensions: mask realism and image-mask consistency.
The researchers evaluated MID on 13 public datasets, covering three modalities: CT, MRI and MRA.
After applying anatomically plausible perturbations across 56 organ groups, MID was able to track the trends of mean Dice and Hausdorff distance without using per-sample ground-truth pairing.
MID can identify distributional shift between different modalities and between different anatomical structures.
Through controlled shuffling experiments (randomly shuffling the pairing of images and masks), MID successfully detected image-mask misalignment.
On multiple real segmentation models, the model rankings produced by MID are reliable, indicating that it is not only a diagnostic tool but can also be used for model selection.
The datasets on which MID has currently been validated are mainly CT, MRI and MRA; whether it is applicable to other medical imaging modalities such as ultrasound and pathology slides is not addressed in the paper.
1 more timeline events
All Facts (11)
Distribution-level metrics have an implicit dependence on the representativeness and size of the reference set—if the reference distribution itself is biased, the evaluation conclusions will shift accordingly.
50%UnverifiedResearchers at Caravel Lab proposed MID (Mask-Image Distributional Divergence), a distribution-level evaluation metric for medical image segmentation.
50%UnverifiedThe starting judgment of MID is: segmentation quality is a property of the joint image-mask distribution.
50%UnverifiedMID directly compares two joint distributions, the reference set (real image-mask pairs) and the prediction set (image-predicted mask pairs), requires no sample-level correspondence at all, and simultaneously captures two dimensions: mask realism and image-mask consistency.
50%UnverifiedThe researchers evaluated MID on 13 public datasets, covering three modalities: CT, MRI and MRA.
50%UnverifiedAfter applying anatomically plausible perturbations across 56 organ groups, MID was able to track the trends of mean Dice and Hausdorff distance without using per-sample ground-truth pairing.
50%UnverifiedMID can identify distributional shift between different modalities and between different anatomical structures.
50%UnverifiedThrough controlled shuffling experiments (randomly shuffling the pairing of images and masks), MID successfully detected image-mask misalignment.
50%UnverifiedOn multiple real segmentation models, the model rankings produced by MID are reliable, indicating that it is not only a diagnostic tool but can also be used for model selection.
50%UnverifiedThe datasets on which MID has currently been validated are mainly CT, MRI and MRA; whether it is applicable to other medical imaging modalities such as ultrasound and pathology slides is not addressed in the paper.
50%UnverifiedThe official implementation of MID has been open-sourced at github.com/erc-caravel/pytorch-mid.
50%