ProstNFound+: A Prospective Study using Medical Foundation Models for Prostate Cancer Detection

Area of Science:

• Artificial Intelligence in Medical Diagnostics
• Medical Imaging Analysis
• Prostate Cancer Detection

Background:

• Medical Foundation Models (FMs) are emerging as powerful tools for developing high-performance diagnostic systems.
• The clinical application of FMs for prostate cancer (PCa) detection using micro-ultrasound (μUS) has not yet been prospectively validated.
• Existing diagnostic protocols for PCa can be complex and rely heavily on expert interpretation.

Purpose of the Study:

• To present ProstNFound+, an adapted foundation model for PCa detection from μUS.
• To conduct the first prospective clinical validation of an FM-based system for PCa detection using μUS.
• To evaluate the model's performance against established clinical scoring systems.

Main Methods:

• ProstNFound+ was developed using a medical FM with adapter tuning and a custom prompt encoder integrating PCa biomarkers.
• The model was trained on retrospective, multicenter μUS data and prospectively evaluated at a new clinical site.
• Model predictions, including cancer heatmaps and risk scores, were benchmarked against PRI-MUS and PI-RADS clinical protocols.

Main Results:

• ProstNFound+ demonstrated strong generalization capabilities, with no performance degradation on prospective data compared to retrospective evaluation.
• The model's predictions closely aligned with standard clinical scores (PRI-MUS and PI-RADS).
• Interpretable heatmaps generated by ProstNFound+ corresponded accurately with biopsy-confirmed PCa lesions.

Conclusions:

• ProstNFound+ shows significant potential for clinical deployment in prostate cancer detection.
• The AI model offers a scalable and interpretable alternative to traditional expert-driven diagnostic protocols.
• Prospective validation confirms the model's reliability and generalizability in a clinical setting.