Predicting clinical endpoints and visual changes with quality-weighted tissue-based renal histological features.

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Data-driven algorithms in digital histopathology face challenges from limited expert annotations and dataset diversity.
• Multi-instance learning (MIL) addresses annotation scarcity in whole slide images (WSI) but often underperforms full supervision.

Purpose of the Study:

• To enhance the effectiveness and scalability of MIL in digital histopathology by incorporating weak annotations.
• To develop an analysis framework for renal biopsy slides (PAS and SR) to predict clinical outcomes.

Main Methods:

• A novel framework was developed to segment renal tissues and extract combined handcrafted and deep features.
• A soft attention model integrated these features for predicting slide-level labels like delayed graft function (DGF) and acute tubular injury (ATI).
• A tissue segmentation quality metric was introduced to mitigate segmentation errors, and the model was trained using 5-fold cross-validation.

Main Results:

• The soft attention model achieved an average ROC-AUC of $$, outperforming ResNet50 ($$), handcrafted features ($$), and baseline methods ($$).
• Weighting tissues by segmentation quality further improved performance by $$.
• The approach demonstrated its utility in supporting clinical decisions through intuitive visualizations pinpointing relevant tissues.

Conclusions:

• Integrating weak annotations significantly boosts MIL performance in digital histopathology, offering a scalable solution.
• The developed framework and attention mechanism provide accurate predictions for renal biopsy classification.
• The method's ability to visualize contributing tissue regions supports clinical decision-making in pathology.