Abstract: TH-PO570
A Computational Pipeline for Segmentation and Classification of Tubules
Session Information
- Pathology and Lab Medicine
November 03, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1700 Pathology and Lab Medicine
Authors
- Sarder, Pinaki, University of Buffalo, Buffalo, New York, United States
- Ginley, Brandon, University of Buffalo, Buffalo, New York, United States
- Lucarelli, Nicholas, University of Buffalo, Buffalo, New York, United States
- Chen, Yijiang, Case Western Reserve University, Cleveland, Ohio, United States
- Hodgin, Jeffrey B., University of Michigan, Ann Arbor, Michigan, United States
- Rosenberg, Avi Z., Johns Hopkins University, Baltimore, Maryland, United States
- Alpers, Charles E., University of Washington, Seattle, Washington, United States
- Madabhushi, Anant, Case Western Reserve University, Cleveland, Ohio, United States
- Barisoni, Laura, Duke Medicine, Durham, North Carolina, United States
- Balis, Ulysses G. J., University of Michigan, Ann Arbor, Michigan, United States
Group or Team Name
- KPMP
Background
The highly repetitive kidney structure is well-suited for high-throughput segmentation using unsupervised methods. Herein, we combined image-based, machine learning (ML) tools to realize an computational pipeline for image curation, segmentation, and classification.
Methods
25 chronic kidney disease and 10 healthy reference tissue were first curated using HistoQC, an open source tool for qualification of whole slide images (WSIs), for histology and imaging artifacts in WSIs that would interfere with ML techniques. WSIs were then processed by the Human-AI-Loop pipeline, a deep-learning-based supervised image segmentation tool, to generate binary masks of all tubules. Segmented tubules were then extracted and processed at the pixel level by the spatially invariant vector quantization (SIVQ) algorithm, which is prepackaged as the validated identification of pre-qualified regions algorithm (VIPR). SIVQ mines a composite vector of biological content inherent in single pixel domains by extracting local kernel goodness-of-fit to a library of pre-selected Fourier signatures of histological primitives. Renal pathologists manually labeled each segmented tubules as normal and abnormal proximal, normal and abnormal distal, and abnormal indeterminate.
Results
3 machine learning-based tubular classes (1-3) were identified by VIPR. There was >95% correlation between manual scoring of proximal tubules (normal and abnormal) and class 1, manual scoring of distal tubules (normal and abnormal) vs. class 2, and for normal vs. abnormal tubular morphology.
Conclusion
Our tools will enable development of large scale feature extraction and statistical quantification of different sub-classes of tubules from giga-pixel size kidney WSIs. This pipeline unleashes the power of artificial intelligence in precision nephrology with the promise of deriving novel digital image biomarkers that can potentially inform disease progression or alignment with molecular markers for theragnostic discoveries.
Funding
- NIDDK Support