Related Experiment Video
Updated: Jul 6, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
An image inpainting-based data augmentation method for improved sclerosed glomerular identification performance with
Songping He1, Yi Zou2, Bin Li1
1Digital Manufacturing Equipment National Engineering Research Center, Huazhong University of Science and Technology, Wuhan, China.
Scientific Reports
|January 10, 2024
Summary
Deep learning models can now quantify glomerulosclerosis in kidney transplants more accurately. By generating synthetic sclerotic glomeruli images, researchers improved segmentation performance, potentially reducing donor kidney discard rates.
Area of Science:
- Nephrology
- Medical Imaging
- Artificial Intelligence
Background:
- Percent global glomerulosclerosis is crucial for renal transplant outcomes.
- Current manual quantification by pathologists is labor-intensive and lacks standardization.
- Deep learning (DL) offers potential for automated and standardized glomerulosclerosis assessment.
Purpose of the Study:
- To improve the identification and segmentation of sclerosed glomeruli using DL.
- To enhance the quantification of percent global glomerulosclerosis for reducing donor kidney discard.
- To address the challenge of imbalanced data (fewer sclerosed than normal glomeruli) in DL models.
Main Methods:
- Utilized 51 publicly available whole slide images (WSIs) from diverse institutions.
- Modified and trained a generative adversarial network (GAN)-based image inpainting model to synthesize sclerosed glomeruli.
- Integrated synthetic sclerosed glomeruli images with real data to train a modified U-Net model for segmentation.
Main Results:
- The proposed inpainting method achieved an average Structural Similarity (SSIM) of 0.8086 and Peak Signal-to-Noise Ratio (PSNR) of 22.8943 dB for generated sclerosed glomeruli.
- Incorporating synthetic images led to improved sclerosed glomerular segmentation performance.
- The modified U-Net model achieved optimal Dice scores for glomerular segmentation across test sets.
Conclusions:
- DL, augmented with GAN-based synthetic data generation, can effectively improve sclerosed glomeruli segmentation.
- This approach enhances the accuracy of percent global glomerulosclerosis quantification.
- The method shows promise for reducing donor kidney discard rates in renal transplantation.
