Beyond VI-RADS Uncertainty: Leveraging Spatiotemporal DCE-MRI to Predict Bladder Cancer Muscle Invasion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• The Vesical Imaging-Reporting and Data System (VI-RADS) shows limited accuracy for non-muscle-invasive bladder cancer (NMIBC) in VI-RADS categories 2 and 3.
• Dynamic contrast-enhanced MRI (DCE-MRI) offers potential for NMIBC assessment by analyzing tumor vascularity but is underutilized due to complex spatiotemporal data.

Purpose of the Study:

• To develop and validate a deep learning model for improved NMIBC diagnosis in VI-RADS 2/3 cases.
• To leverage spatiotemporal features from DCE-MRI for enhanced bladder cancer classification.

Main Methods:

• A deep learning model was created to quantify spatiotemporal features from multiphase DCE-MRI in 184 patients (VI-RADS 2 or 3).
• The model incorporated multiscale feature extraction and contextual attention mechanisms.
• Model performance was evaluated against benchmarks and conventional thresholds.

Main Results:

• The deep learning model achieved a sensitivity of 0.90 for NMIBC and an AUC of 0.82 for overall classification in VI-RADS 2/3 cases.
• It outperformed existing benchmarks and the VI-RADS ≤ 2 threshold (sensitivity 0.67).
• Visualizations confirmed the model's ability to identify key spatiotemporal patterns related to muscle invasion.

Conclusions:

• Deep learning analysis of DCE-MRI spatiotemporal data significantly enhances NMIBC diagnosis accuracy for VI-RADS 2/3 classifications.
• This AI-driven approach provides a valuable tool to overcome current VI-RADS assessment limitations.
• The model offers improved clinical decision-making for bladder cancer staging.