Abstract: TH-PO029
Developing a Novel Deep Learning-Based Method for Kidney Transmission Electron Microscopy Image Analysis
Session Information
- Augmented Intelligence for Prediction and Image Analysis
October 24, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Zou, Anqi, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Ji, Jiayi, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Fan, Xueping, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Tan, Winston, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Dodd, Laura, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Oei, Emily, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Chen, Hui, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Liu, Yu-Chen, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Lu, Simon L., Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Henderson, Joel M., Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Zhang, Chao, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
- Lu, Weining, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States
Background
Abnormal glomerular basement membrane (GBM) and podocyte foot process (PFP) widths are critical diagnostic criteria for proteinuric kidney disease. The standard method for quantifying GBM and PFP widths is unbiased stereology, which is labor-intensive and not routinely used in research or clinical diagnosis. To address this gap, we developed a novel deep learning (DL) based algorithm to quantify GBM and PFP widths from glomerular transmission electron microscopy (TEM) images.
Methods
Our algorithm has a two-stage workflow (see Figure): (a) We first used DL for image segmentation, employing a PSPNet-based model with self-training techniques to iteratively collect annotated data, reducing the manual labeling time and improving accuracy. (b) We then used image processing techniques to measure GBM and PFP widths, developing a graph-based method to refine GBM skeleton extraction and compute GBM and PFP widths. We used glomerular TEM datasets from two animal models of proteinuric kidney disease to train and validate our algorithm: the ILK podocyte-specific knockout (ILK cKO) mice and the Passive Heymann’s Nephritis (PHN) rats.
Results
The segmentation performance on the validation dataset showed a mean IoU of 0.77, Dice of 0.87, and accuracy of 0.88. In a PHN rat validation dataset, the relative mean error compared to manual measurements is 0.7% for GBM width and 3% for PFP width, indicating overall interchangeability between these two methods. Comparisons of DL-based measurements between wild-type (WT) and ILK cKO groups showed significant differences in GBM width (p = 1.5e-07) and PFP width (p = 3.1e-06), suggesting that the method effectively distinguished healthy from pathological specimens.
Conclusion
We have established and validated a new DL-based algorithm for automatic GBM and PFP width measurements on rodent glomerular TEM images. This tool is compatible with human kidney biopsy TEM images, which can be translated into clinical applications.
Two-stage workflow to quantify kidney GBM and PFP widths: (a) DL for image segmentation. (b) Image processing techniques to measure GBM and PFP widths.
Funding
- NIDDK Support