Related Experiment Video
Updated: Feb 16, 2026
10:39
A Label-Free Segmentation Approach for Intravital Imaging of Mammary Tumor Microenvironment
Published on: May 24, 2022
2.8K
The first MICCAI challenge on PET tumor segmentation.
Mathieu Hatt1, Baptiste Laurent1, Anouar Ouahabi1
1LaTIM, UMR 1101, INSERM, IBSAM, UBO, UBL, Brest, France.
Medical Image Analysis
|December 22, 2017
Summary
A benchmark dataset and challenge were created for comparing automatic PET image segmentation methods. Convolutional neural networks showed strong performance, though not universally superior to traditional algorithms.
Area of Science:
- Medical Imaging
- Image Analysis
- Computational Biology
Background:
- Automatic functional volume segmentation in PET images is complex, with numerous methods lacking standardized comparison datasets.
- The absence of a benchmark dataset hinders direct evaluation and comparison of segmentation algorithm performance across studies.
Purpose of the Study:
- To establish a benchmark dataset and conduct a comparative study of recent PET image segmentation methods.
- To facilitate direct comparison of segmentation algorithm results following established guidelines (AAPM TG 211).
Main Methods:
- Assembled a large dataset (176 images) including simulated, phantom, and clinical PET images.
- Organized a challenge where participants submitted encapsulated pipelines for autonomous algorithm execution and evaluation.
- Ranked methods based on the arithmetic mean of sensitivity and positive predictive value.
Main Results:
- Ten methods from four teams were evaluated, alongside two thresholding methods and Fuzzy Locally Adaptive Bayesian (FLAB).
- All but one method achieved median accuracy above 0.8.
- Convolutional neural network-based segmentation achieved the highest score, outperforming 9 out of 12 other methods.
Conclusions:
- This study represents the most rigorous comparative analysis of PET segmentation algorithms to date.
- Method performance hierarchy was consistent across different datasets and metrics.
- The majority of submitted methods demonstrated robust performance with median accuracy exceeding 0.8.
