Parameter-Efficient Fine-Tuning of MedSAM for Prostate and Urethra Segmentation in Brachytherapy TRUS Images

Rucha Bhalchandra Joshi, Eleftherios Papageorgiou, Myrianthi Hadjicharalambous, Yiannis Roussakis, Georgios Anagnostopoulos, Iosif Strouthos, Constantinos Zamboglou, Constantine Dovrolis

Keywords: Parameter-efficient fine-tuning, MedSAM, Prostate brachytherapy, TRUS image segmentation, Foundation models for medical imaging

TL;DR: Parameter-efficient MedSAM fine-tuning enables fast, highly accurate prostate and urethra segmentation in TRUS images, reducing brachytherapy workflow time from minutes to seconds.

Abstract:

Prostate and urethra segmentation in transrectal ultrasound (TRUS) images during brachytherapy is commonly performed manually, a process that is time-consuming, often exceeding 20 minutes - particularly when metallic brachytherapy needles introduce artifacts that obscure organ boundaries. The prolonged operating room time adds to staff burden and patient discomfort under anesthesia. Automated segmentation using medical foundation models such as MedSAM offers a direct solution, by reducing procedure time and improving brachytherapy workflow efficiency. Towards that aim, in this study we systematically evaluate 10 parameter-efficient fine-tuning strategies for MedSAM on 204 TRUS volumes, containing needle artifacts from brachytherapy procedures. The variants ranged from full retraining (100% parameters) to lightweight LoRA-based adaptations (<1% parameters), targeting different architectural components of MedSAM.

The best-performing variant, MD Transformer LoRA, achieved a volume Dice score of $0.9484$ [95% CI $0.9449, 0.9517$] for prostate segmentation and $0.9807$ [95% CI $0.9800, 0.9813$] for urethra segmentation while training only 0.09% of model parameters. Parameter efficient variants consistently matched or exceeded full retraining performance across both in-house and three external datasets (3, 11, and 72 patients), substantially outperforming the original pretrained MedSAM (Dice: $0.8147$) and nnU-Net baseline $0.8917$. Based on our results, automated segmentation using parameter-efficient MedSAM fine-tuning can reliably replace manual delineation, reducing operation room time for the segmentation task from $20+$ minutes to around 6 seconds per patient. This approach enables clinical deployment with minimal computational overhead while maintaining high accuracy (even in the presence of needle-induced artifacts), ultimately improving brachytherapy workflow efficiency and patient care.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Reproducibility: https://github.com/ruchajoshi/PROTECT

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 31