MedScribble Multi-image Segmentation Biomedical Task Dataset

*This dataset supports online use.Click here to jump.

This dataset was released by a research team from MIT in 2024. The related paper is “ScribblePrompt: Fast and Flexible Interactive Segmentation for Any Biomedical Image", which has been accepted by ECCV24.

The dataset is handwritten scribbles collected by the research team from 14 different open-access biomedical image segmentation datasets for 14 segmentation tasks by 3 annotators. MedScribble contains a total of 64 pairs of 2D image segmentation pairs, each with 3 sets of scribble annotations.

For each segmentation task (i.e., dataset/label combination), annotators were shown 5 training examples with true segmentations and were instructed to draw positive and negative scribbles on new images to indicate regions of interest.

Annotators draw doodles in a Gradio web app. Annotators 1 and 2 use iPads with styluses, while Annotator 3 uses a laptop trackpad to draw doodles.

All images were padded to squares (with zeros) before being resized to 256×256 and rescaled to the [0,1] range. For the 3D dataset, the research team selected the middle slice (midslice) or the slice with the largest label area (maxslice) as indicated by the folder name.

HyperAI Super Neural Paper Interpretation:Selected for ECCV 2024! Covering 54,000+ images, MIT proposed a general model for medical image segmentation, ScribblePrompt, which performs better than SAM"

Try the tutorial demo now:ScribblePrompt Medical Image Segmentation Tool