ProMUS-NET: Artificial intelligence detects more prostate cancer than urologists on micro-ultrasonography

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Urology

Background:

• Prostate cancer detection relies on imaging and biopsy.
• Micro-ultrasonography (MUS) offers high-resolution visualization.
• Improving accuracy and consistency in cancer localization is crucial.

Purpose of the Study:

• Develop a deep learning model for automatic prostate cancer segmentation on MUS.
• Enhance sensitivity and inter-reader consistency in cancer localization.
• Compare the AI model's performance against expert urologists.

Main Methods:

• Prospective collection of MUS images from patients undergoing MRI-ultrasonography fusion biopsy.
• Annotation of clinically significant cancer (Grade Group ≥2) on MUS images.
• Training a U-Net based model (ProMUS-NET) using fivefold cross-validation.

Main Results:

• The AI model achieved an Area Under the Curve (AUC) of 0.92.
• ProMUS-NET demonstrated higher lesion-level (73% vs 58%) and patient-level (77% vs 66%) sensitivity compared to urologists.
• High sensitivity (86.2%) was observed for peripheral zone lesions.

Conclusions:

• The AI model accurately identifies prostate cancer lesions on MUS with high sensitivity.
• AI-assisted detection shows potential to improve biopsy diagnosis accuracy.
• Further research will focus on external validation and reducing false positives.