Micro-ultrasound (MicroUS) is an emerging high-resolution imaging modality for prostate cancer diagnosis and real-time biopsy guidance. Automated MicroUS analysis is constrained by the need for extensive expert annotation. The authors developed a modality-specific foundation model, muFM, to evaluate whether self-supervised pretraining on large unlabeled MicroUS datasets can enable data-efficient downstream tasks including segmentation and lesion localization.
The foundation model muFM was pretrained using a DINOv2-style self-supervised learning approach on a dataset of 1,736,391 unlabeled MicroUS images collected from 780 patients. After pretraining, muFM was fine-tuned and evaluated on three downstream cohorts representing distinct clinical tasks:
The authors compared muFM to several alternatives: randomly initialized models and foundation models pretrained on natural images, other medical images, or non–MicroUS ultrasound images. Fine-tuning experiments included scenarios with limited annotated data to assess data efficiency.
Segmentation performance was measured with Dice scores and the 95th-percentile Hausdorff distance (HD95). Lesion localization performance was assessed using free-response receiver operating characteristic (FROC) analysis, reporting sensitivity at specified false-positive detection rates per 3D volume. Comparisons focused on performance under varying annotation budgets, highlighting gains when annotated data were scarce.
muFM improved prostate capsule and urethra segmentation across evaluated datasets, with the largest gains when annotation was limited. Reported results using 10% of training scans included:
Prostate capsule segmentation: muFM achieved a Dice score of 93.8%, compared with 87.9% for the strongest comparator model. The corresponding HD95 was 2.0 mm for muFM.
Urethra segmentation: muFM achieved a Dice score of 58.4%, compared with 47.0% for the strongest comparator model. The corresponding HD95 was 4.3 mm for muFM.
These results indicate that MicroUS-specific pretraining materially improved segmentation quality with substantially reduced annotation requirements relative to other initialization strategies.
For prostate cancer lesion localization, muFM outperformed other pretrained models in FROC analysis. The study reports that muFM attained approximately 0.75 sensitivity at two false-positive detections per 3D volume, exceeding the sensitivity achieved by comparator models under similar conditions. This suggests improved detection performance for lesion localization when leveraging modality-specific self-supervised pretraining.
The authors reported several limitations in the study design and scope:
Where details were unspecified in the source, such as exactly which comparator models and their pretraining corpora were used beyond the broad categories reported, those specifics were not provided in the article text.
The study demonstrates that a MicroUS-specific, self-supervised foundation model (muFM) pretrained on a large corpus of unlabeled MicroUS images can substantially improve data-efficient image analysis for prostate applications. Improvements were observed for prostate capsule segmentation, urethra segmentation, and prostate cancer lesion localization, particularly when annotated training data were limited. The authors propose that modality-specific self-supervised pretraining may reduce annotation burden when developing automated MicroUS tools and could accelerate deployment of clinical algorithms for real-time biopsy guidance and diagnostic support.
Ethical approval was granted by the Institutional Review Board of the University of California, Los Angeles, which waived the requirement for informed consent. The data are not publicly available due to privacy and institutional restrictions but may be requested from the corresponding author with IRB permission.
Funding sources declared include NIH grants R01-CA248506 and R01-CA272702. Competing interests disclosed by the authors include a financial interest in Siemens Healthineers by one author, research funding and honoraria from Exact Imaging to another author, consultancy for Exact Imaging to a third author, and affiliation of one author with Exact Imaging, the manufacturer of the ExactVu MicroUS system used in the study. Remaining authors reported no competing interests.