Micro-Ultrasound-Specific Adaptation of the Foundation Model for Prostate Segmentation
Abstract
Purpose
Micro-ultrasound (microUS) provides high-resolution visualization of the prostate for interventional procedures; however, the scarcity of annotated datasets limits the development of robust automated segmentation methods. This study leverages a foundation-model–based strategy with self-supervised learning to enhance prostate boundary delineation and reduce manual contouring burden in microUS-guided biopsy and brachytherapy.
Methods
We propose μPro-DINO, which adapts a DINOv3 ViT-B vision foundation model for prostate segmentation on 29-MHz microUS without extensive in-domain pretraining. To address microUS-specific speckle texture and indistinct gland margins, we introduce an Adaptive Texture–Semantics Fusion module with token-wise gating that dynamically integrates shallow, texture-sensitive features with deeper semantic representations. Evaluation was performed on the public Micro-Ultrasound Prostate Segmentation Dataset (75 patients) using the released split (55 training / 20 testing) and subject-level five-fold cross-validation. Clinically relevant segmentation quality was assessed using the Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95), mean surface distance (MSD), and volume difference, with emphasis on boundary accuracy relevant to targeting and applicator guidance.
Results
On the held-out test set, μPro-DINO achieved a DSC of 0.946±0.018, HD95 of 1.84mm, MSD of 0.48mm, and volume difference of 5.05%, outperforming MicroSegNet (DSC 0.937±0.020; HD95 2.26mm; MSD 0.66mm; volume difference 5.49%). Boundary accuracy improved substantially, with HD95 reduced by 0.42mm and MSD reduced by 0.18mm, indicating tighter gland surface agreement in regions where margins are poorly visualized on microUS. Across five-fold cross-validation, μPro-DINO demonstrated consistent performance (DSC 0.950±0.017; HD95 1.68mm; MSD 0.42mm; volume difference 4.39%), confirming stable generalization across patient partitions.
Conclusion
μPro-DINO provides clinically meaningful improvements in prostate boundary precision on micro-ultrasound, with the potential to reduce contour editing and improve consistency for microUS-guided biopsy and brachytherapy. These findings support foundation-model adaptation as a practical and scalable approach for reliable prostate segmentation in interventional and radiation oncology workflows.