Related Experiment Video
Updated: Aug 1, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Impact of imperfection in medical imaging data on deep learning-based segmentation performance: An experimental study
Ayetullah Mehdi Güneş1, Ward van Rooij1, Sadaf Gulshad2
1Department of Radiation Oncology, Amsterdam UMC, Amsterdam, The Netherlands.
Medical Physics
|April 29, 2023
Summary
Deep learning models for parotid gland segmentation require high-quality training data. While imperfect segmentations degrade performance, training with lower-quality imaging data can improve model robustness.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
Background:
- Clinical data for deep learning often contains imperfections in imaging and segmentations.
- These imperfections can impact model performance and reliability.
Purpose of the Study:
- Investigate the effect of data imperfections on deep learning models for parotid gland segmentation.
- Utilize synthesized data for controlled experimentation.
- Enhance the reliability and performance of deep learning models.
Main Methods:
- Synthesized data from clinical segmentations to create pseudo ground-truth.
- Simulated three types of imperfections: incorrect segmentations, low contrast, and artifacts.
- Varied imperfection severity across five levels and cross-evaluated models.
Main Results:
- Synthesized data without errors yielded near-perfect segmentation.
- Reduced segmentation quality in training data significantly decreased model performance.
- Training with lower image quality (reduced contrast, artifacts) improved model robustness.
Conclusions:
- High-quality segmentations are crucial for effective deep learning model training.
- Training deep learning models with imperfect imaging data can enhance their robustness to such imperfections.
