Related Experiment Video
Updated: Jan 13, 2026
Supervised Machine Learning for Semi-Quantification of Extracellular DNA in Glomerulonephritis
Published on: June 18, 2020
SegRenal: AI-Driven segmentation of frozen sections in transplant kidney biopsies - A comparative analysis of deep
Ibrahim Yilmaz1, Heba M Alazab1, Fatih Doganay1
1Department of Laboratory Medicine and Pathology, Mayo Clinic, Jacksonville, 32224, FL, USA.
Background And Objective:
Frozen section evaluation of donor kidney biopsies is vital for determining transplant suitability, yet remains challenging due to interobserver variability and freezing-related artifacts. While deep learning (DL) has been used for permanent sections, its application to frozen tissue is limited. We developed SegRenal, an artificial intelligence (AI)-based segmentation model for automated identification of glomeruli (non-sclerotic and sclerotic), arteries, and interstitial fibrosis and tubular atrophy (IFTA) in hematoxylin and eosin-stained frozen whole-slide images (WSIs). This study focuses on rigorous model adaptation, dataset development, cross-scanner performance evaluation, and integration into a clinical digital pathology workflow.
Methods:
A total of 183 frozen WSIs were collected from two scanners (GT450 and Grundium) and manually annotated by expert renal pathologists. Three encoder-decoder architectures (UNet, ResNet-UNet, and DenseNet-UNet) were trained on patch-level data to compare binary and multiclass segmentation strategies. The best-performing configuration - pre-trained DenseNet with multiclass output - was further evaluated on 21 unseen WSIs scanned with both platforms. Preprocessing included downsampling, patch extraction, and annotation refinement. Performance was assessed using Dice score, precision, recall, and intraclass correlation coefficient (ICC). Bland-Altman analysis and scanner variability experiments were conducted.
Results:
Pre-trained DenseNet multiclass segmentation yielded the best overall performance: Dice scores of 0.95 (glomeruli), 0.90 (sclerotic glomeruli), 0.80 (arteries), and 0.88 (IFTA). Recall reached 99.8% for glomeruli and 100% for arteries. Performance remained consistent across scanners. In several cases, the model detected structures initially missed by manual annotation, later confirmed by pathologists.
Conclusions:
SegRenal accurately segments key renal compartments in frozen biopsies and demonstrates robust cross-scanner performance. By automating tissue quantification, the model reduces variability and turnaround time, supporting fast and consistent intraoperative kidney transplant assessments.
