MUSeg-PSV: A Real-Time Deep-Learning Segmentation of the Prostate Gland and Seminal Vesicles on 29-MHz...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Prostate cancer diagnosis and staging rely on accurate imaging.
• Micro-ultrasound (MicroUS) offers high resolution but requires precise anatomical guidance.
• Segmentation of prostate and seminal vesicles is crucial for biopsy and staging accuracy.

Purpose of the Study:

• To develop and evaluate MUSeg-PSV, a deep learning framework for real-time segmentation of the prostate and seminal vesicles.
• To enhance anatomical orientation during MicroUS-guided prostate biopsy.
• To support local staging of prostate cancer using AI-assisted MicroUS.

Main Methods:

• A lightweight, single-stage convolutional neural network (YOLOv11s-seg) was adapted for segmentation.
• Two training strategies were compared: standard augmentation and augmentation-rich configurations.
• Performance was evaluated using metrics like mAP, DSC, ECE, and Brier score, alongside inference latency.

Main Results:

• The augmentation-rich model significantly improved mean average precision (mAP) and prostate segmentation (DSC).
• Seminal vesicle detection showed substantial improvement (AP: 0.183 to 0.842) with robust DSC.
• Improved calibration (ECE, Brier score) and real-time performance (28-35 ms/frame) were achieved.

Conclusions:

• MUSeg-PSV provides reliable, real-time segmentation of prostate and seminal vesicles on MicroUS.
• The AI framework enhances operator orientation for biopsies and prostate cancer staging.
• This technology promotes standardized identification of seminal vesicles in clinical workflows.