Abstract: FR-OR015
TRIMS: Deep-Learning Segmentation of Renal Tubules and Immune Cells in 32-Marker Spatial Proteomic Images from Lupus Nephritis Biopsies
Session Information
- Artificial Intelligence and Data Science Transforming Kidney Care: From Algorithms to Action
November 07, 2025 | Location: Room 361A, Convention Center
Abstract Time: 05:20 PM - 05:30 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Gottipati, Mahith, Dulles High School, Sugar Land, Texas, United States
- Kommisetty, Suchay, Jordan High School, Fulshear, Texas, United States
- Daouk, Mohammad, University of Houston, Houston, Texas, United States
- Mohan, Chandra, University of Houston, Houston, Texas, United States
- Louis Sam Titus, Anto Sam Crosslee, University of Houston, Houston, Texas, United States
Background
Lupus nephritis (LN) injures renal tubules as well as glomeruli, yet tubulointerstitial injury and immune-cell infiltration are still graded manually and subjectively.
Methods
Kidney tissue from six class III–V LN patients and six controls was stained with a 32-plex Phenocycler™ panel (tubular markers β-catenin/E-cadherin/pan-cytokeratin/AQP3; immune markers CD45/CD3ε/CD4/CD8). From the Whole Slide Images (WSIs) we polygon-annotated 298 tubule tiles (200×200 px) and 324 immune-cell tiles (40×40 px) in QuPath. Rotational/flip augmentation enlarged the set to 1784 tubule tiles/1944 immune-cell tiles. Annotations were converted to COCO format, corrected, and split into 70/15/15 training/validation/test sets. Tubule & Renal Immune-cell Models for Segmentation (TRIMS), 2 Detectron2 Mask R-CNN pipelines, were developed and trained on an A100 GPU.
Results
In held-out test sets TRIMS achieved an F1-score=0.90/0.86, precision=0.88/0.88, and recall=0.94/0.85 (Fig 1.). Visual inspection confirmed accurate segmentation (including avoiding glomeruli) in dense and heterogeneous regions.
Conclusion
TRIMS accurately segmented renal tubules and immune-cells, offering a consistent and objective alternative to traditional DAPI-based and manual methods, and exceeded the performance of competing platforms. TRIMS provides fast, objective, morphology-aware segmentation of renal tubules and immune infiltrates, enabling large-scale clinicopathologic and spatial-omics studies.
Figure 1. Model predictions/metrics at varying epochs. EN=Epoch Number, LR=Learning Rate, F1=F1-Score. A. Predictions of models trained at 6 epoch counts. Region coverage improves with training. B. Performance metrics at 6 epoch counts. For immune-cells, the improved results across extended training correlate with expectation. The tubule model peaked at 15000 epochs, with a dip at 12000 indicating brief overfitting.