Evaluating Medical Image Segmentation Models Using Augmentation
Author Information
Author(s): Sayed Mattin, Saba-Sadiya Sari, Wichtlhuber Benedikt, Dietz Julia, Neitzel Matthias, Keller Leopold, Roig Gemma, Bucher Andreas M.
Primary Institution: Goethe University Frankfurt
Hypothesis
Can a novel validation framework using data augmentation effectively assess the accuracy of automated medical image segmentation models?
Conclusion
The developed framework allows for evaluating segmentation performance without needing manually labeled ground truth data.
Supporting Evidence
- The study demonstrated strong correlation between segmentation quality of original scans and alignment metrics of augmented scans.
- Average DICE scores indicated high agreement between ground truth and model-generated segmentations.
- Uncertainty heatmaps provided insights into model performance and areas needing improvement.
Takeaway
This study created a new way to check if computer programs that help doctors see inside the body are doing a good job, without needing a lot of extra work.
Methodology
The study used 101 CT scans and their corresponding segmentation masks, applying data augmentation to assess model consistency and performance metrics.
Potential Biases
Potential bias due to using the same dataset for both training and evaluation.
Limitations
The evaluation metrics may be biased as the same dataset was used for training and testing, and the sample size was limited to 101 CT scans.
Statistical Information
P-Value
p<0.01
Statistical Significance
p<0.01
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website