Explainable Computational Imaging for Precision Oncology: An Interpretable Deep Learning Framework for Bladder Cancer...

Area of Science:

• * Computational pathology
• * Artificial intelligence in oncology
• * Medical image analysis

Background:

• * Bladder cancer is a global health concern with high recurrence rates.
• * Current diagnostic methods are invasive, time-consuming, and subjective.
• * There is a need for accurate, efficient, and objective diagnostic tools.

Purpose of the Study:

• * To develop and evaluate a transparent deep learning model for bladder cancer detection.
• * To enhance the accuracy and efficiency of histopathology slide analysis.
• * To provide visual support for AI-driven diagnostic predictions.

Main Methods:

• * Training and testing five variants of the YOLOv11 deep learning architecture (nano, small, medium, large, extra large).
• * Utilizing a dataset of hematoxylin and eosin-stained histopathology slides categorized into inflammation, urothelial cell carcinoma (UCC), and invalid tissue.
• * Employing performance metrics including accuracy, precision, recall, AUPRC, ROC-AUC, risk-coverage analysis, and Expected Calibration Error (ECE).

Main Results:

• * The YOLOv11-large model achieved the highest performance with 97.09% accuracy, 95.47% precision, and 95.47% recall.
• * Detailed assessments using AUPRC, ROC-AUC, and risk-coverage analysis confirmed the model's steadiness and trustworthiness.
• * The model demonstrated excellent calibration (low ECE) and identified potential morphological overlap between inflammation and invalid samples.

Conclusions:

• * The YOLOv11-large model represents a significant advancement in AI-assisted bladder cancer diagnosis.
• * The transparent nature of the model provides visual support, enhancing reliability and interpretability.
• * This computationally efficient and scalable AI system moves closer to practical clinical application.