Prostate micro-ultrasound: what µFM, a foundation model pretrained on 1.7 million images, changes when annotations are scarce
A UCLA team trained µFM, a foundation model (a large model pretrained without labels, then adapted to specific tasks), on 1,736,391 prostate micro-ultrasound images, a very high-resolution ultrasound used to guide biopsies. With only 10% of the scans annotated, µFM reaches a Dice of 93.8% on the prostate capsule and 58.4% on the urethra, versus 87.9% and 47.0% for the strongest comparator, and detects clinically significant lesions with a sensitivity of 0.762 at two false alerts per volume. The advantage is clear when annotations are scarce and fades when everything is annotated; testing remains single-centre and retrospective, with no external validation for two of three tasks, and several authors have ties to the device manufacturer.
The context
Micro-ultrasound (“microUS”) is a high-resolution imaging modality that the authors describe as emerging for prostate cancer diagnosis and real-time biopsy guidance. Reading it remains difficult: one must outline the prostate, locate the urethra to spare it during focal treatment, and decide which areas to target. Each task normally requires many images annotated by hand by experts, which is slow and costly.
Foundation models try to bypass this bottleneck: a large network is pretrained on a mass of images without labels so that it learns the visual regularities of the domain, then adjusted (fine-tuning) with few annotated examples. Several such models already exist for ultrasound in general. The bet of this paper is that a model specific to one modality, microUS, should outperform generalist or “all-ultrasound” models.
The method
The model. µFM is built on a ViT-B/16, a Vision Transformer, an architecture that cuts the image into 16 × 16 pixel patches treated like the words of a sentence. Its pretraining follows the DINOv2 recipe: a “student” network must reproduce the representations of a “teacher” network from distorted or masked views of the same image (DINO and iBOT objectives), with no labels. The paper does not give the parameter count.
Pretraining data. 1,736,391 images from 6,610 scans of 780 patients, acquired at UCLA between January 2023 and April 2026, before and after biopsy, all on a single microUS platform with a single operator. Patients later used for testing are excluded from pretraining.
The three downstream tasks. Prostate capsule segmentation (public University of Florida dataset: 75 patients, split 45 train, 10 validation, 20 test); urethra segmentation (UCLA: 105 patients, 60/15/30); and localization of clinically significant lesions, meaning grade group 2 or higher (UCLA: 190 patients, 110/22/58, with 48 lesion-positive and 10 lesion-negative patients in the test set). Segmentation means assigning each pixel a label (“capsule”, “urethra”, “background”).
Comparators. The same ViT-B randomly initialized, or pretrained with iBOT and DINOv2 (natural images), BiomedCLIP (biomedical images) and four ultrasound models (USFM, URFM, USF-MAE, EchoCare), plus three networks trained from scratch: nnU-Net (a segmentation framework that configures its own pipeline), TransUNet and MicroSegNet. Segmentation models are fine-tuned on 10, 20, 50 and 100% of the training scans, with five repetitions using different random seeds. Statistics rely on Wilcoxon tests with Holm correction.
Metrics. Dice measures overlap between predicted and reference masks (1 = perfect). HD95 is the distance, in millimetres, within which 95% of the predicted contour lies relative to the reference contour: lower is better. FROC plots sensitivity (share of lesions found) against the number of false detections per volume.
The results
ML metrics. Capsule, with 10% of scans: Dice 93.8 ± 0.6% and HD95 2.0 ± 0.4 mm for µFM, versus 87.9% and 5.8 mm for EchoCare, the best pretrained comparator, and 85.8% for nnU-Net. Urethra, with 10%: Dice 58.4 ± 1.0% and HD95 4.3 mm, versus 47.0% and 6.8 mm for EchoCare. With 100% of scans the gap narrows: 95.1% for µFM versus 94.6% for nnU-Net on the capsule, not significant (pHolm = 0.648 for Dice); 63.1% versus 59.4% for EchoCare and 58.2% for MicroSegNet on the urethra, not significant against these two. For lesions, at one false detection per volume sensitivity is 0.648 ± 0.023 (URFM: 0.507), and at two false detections 0.762 ± 0.030 (URFM: 0.559). No statistical test is reported for this task.
Clinical translation. For the capsule, 10% of training scans means roughly five scans here (computed from the 45 training patients); reaching 93.8% Dice with so little annotation matters to a department that wants to adapt the tool to its own machine. A Dice of 58.4% on the urethra means false detections plus omissions amount to about 1.4 times the correctly segmented volume: the contour stays approximate, which matters for a structure to be spared during focal treatment. For detection, at two false alerts per volume about three lesions in four are found: of 100 real lesions, roughly 24 missed versus roughly 44 for the best comparator, but each exam would on average raise two areas to recheck. The exact number of test lesions is not given in the text.
What is good
Scale and a broad comparison. 1.7 million images pretrained on a narrow domain, against eleven references from four families (natural images, biomedical, general ultrasound, supervised networks), with five seeds and tests corrected for multiple comparisons.
A measure of annotation efficiency. Results are given at 10, 20, 50 and 100% of scans, showing where pretraining helps and where it stops helping. The authors do not hide that at 100%, nnU-Net matches µFM on the capsule.
Careful patient separation. Downstream patients are excluded from pretraining at the patient level, which limits data leakage (information flowing from test to training). The capsule dataset also comes from another institution, the University of Florida, giving a first off-site test for that task.
What is less good
Population and device bias. Pretraining and two of three tasks come from a single hospital, a single platform and a single operator. The risk is shortcut learning: the model may rely on signatures of the device or of that operator's scanning style rather than anatomy. The authors acknowledge the absence of external or prospective validation for the urethra and lesions.
Small test sets and poorly described uncertainty. Test sets contain 20, 30 and 58 patients. The standard deviations cover five training repetitions, not the sampling of patients: they measure optimization stability, not the statistical uncertainty of performance. No test is given for lesion detection, and no ablation isolates what drives the gain: domain specificity, data volume or training recipe. Moreover, µFM is pretrained on images from the institution that also supplies the urethra and lesion test sets, whereas, per the text, the comparators come from other sources: the comparison is less symmetric than it looks.
A flattering metric and a task far from clinical practice. A Dice of 95% on the capsule leaves little room: at 100% annotation the difference from nnU-Net is not significant. For lesions, a sensitivity of 0.762 at two false detections per volume is one point on an averaged FROC curve, not a patient-level detection rate. The paper measures algorithmic performance, not its effect on biopsy targeting or workflow. Add conflicts of interest: one author is affiliated with Exact Imaging, maker of the ExactVu device used, two others receive funding or fees from that company, and another has a financial interest in Siemens Healthineers. Public funding comes from the NIH (R01-CA248506 and R01-CA272702). No statement on code or weights availability is made, and the data are not public.
What it changes
For the research community. The paper provides a solid reference point for microUS and suggests a modality-specific model helps mostly when annotations are scarce. The logical next steps: test it at other centres and on other machines, with ablations, and release code and weights so others can verify it.
For clinicians. Nothing changes today: this is a preprint, not a validated tool, and no result concerns the decision to biopsy. In time, a base model adaptable with a few dozen annotations could lower the cost of setting up segmentation or targeting-aid tools, provided it is shown to hold from one hospital to another.
For patients and the public. Better delineation of the prostate, urethra and suspicious areas could, if validation follows, make biopsies and focal treatments more precise. We are far from that: the paper says nothing about missed diagnoses or biopsies avoided in practice.
Going further
The preprint is available on medRxiv: doi.org/10.64898/2026.09.29.26364330. It relies on the public MicroSegNet dataset (University of Florida) for the prostate capsule. The comparators cited (DINOv2, BiomedCLIP, USFM, URFM, USF-MAE, EchoCare, nnU-Net, TransUNet, MicroSegNet) are described in their respective publications. The preprint has not been peer reviewed; its license is CC BY-NC-ND 4.0.
Editorial transparency: this article was produced with the help of artificial intelligence tools from the preprint text; the English, Spanish and Chinese versions are machine translations.