NMGrad: Advancing Histopathological Bladder Cancer Grading with Weakly Supervised Deep Learning.

Area of Science:

• Oncology
• Digital Pathology
• Machine Learning

Background:

• Urothelial carcinoma is the most common bladder cancer, with high recurrence rates and associated costs.
• Accurate grading is crucial for risk stratification but suffers from pathologist variability.
• Lack of annotations in medical images hinders deep learning model training for bladder cancer grading.

Purpose of the Study:

• To develop and evaluate an automated pipeline for bladder cancer grading using histological slides.
• To address the challenges of grading inconsistencies and the need for annotated data in deep learning.

Main Methods:

• A pipeline was developed involving urothelium tissue tile extraction at multiple magnification levels.
• A convolutional neural network was used for feature extraction from tissue tiles.
• A nested multiple-instance learning approach with attention was employed for slide-level grade prediction, incorporating tile origin for malignancy level distinction.

Main Results:

• The model achieved an F1 score of 0.85, outperforming previous state-of-the-art methods.
• Attention scores at the region level correlated with verified high-grade regions, providing model explainability.
• Clinical evaluations confirmed the model's consistent superior performance.

Conclusions:

• The proposed pipeline offers a robust and explainable solution for automated bladder cancer grading from histological slides.
• This approach can help overcome pathologist variability and improve risk stratification for urothelial carcinoma patients.
• The model's performance suggests significant potential for clinical application in digital pathology workflows.