MRI-Based Bladder Cancer Staging via YOLOv11 Segmentation and Deep Learning Classification.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate bladder cancer staging is crucial for treatment decisions, especially differentiating non-muscle-invasive (T1) from muscle-invasive (T2-T4) disease.
• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast but is limited by operator-dependent interpretation and inter-observer variability.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for standardized MRI-based bladder cancer staging.
• To support reproducible radiological interpretation and improve clinical management decisions.

Main Methods:

• A sequential AI pipeline integrating YOLOv11 for tumor segmentation and DeepLabV3 for boundary refinement was developed.
• Three deep learning classifiers (VGG19, ResNet50, Vision Transformer) were trained on 416 T2-weighted MRI images for stage prediction.
• Performance was assessed using accuracy, precision, recall, F1-score, and multi-class AUC, with uncertainty characterized by bootstrap confidence intervals.

Main Results:

• All evaluated models demonstrated high and comparable discriminative performance for MRI-based bladder cancer staging.
• High accuracy and AUC were achieved, particularly in differentiating non-muscle-invasive from muscle-invasive bladder cancer.
• Calibration analysis confirmed the probabilistic behavior of predicted stage probabilities.

Conclusions:

• The proposed deep learning framework demonstrates the feasibility of automated MRI-based bladder cancer staging.
• The framework supports the potential of AI for standardizing and reproducing MRI-based staging procedures, serving as a methodological tool for consistency.
• Further validation with multi-center datasets, pathology confirmation, and explainable AI is needed for generalizability and clinical relevance.