Summary
A medRxiv preprint reports that muFM, a model pretrained on more than 1.7 million unlabeled micro-ultrasound images, improved performance on three prostate image-analysis tasks. Gains were largest when the model was fine-tuned with limited annotated data.
A preprint describes muFM, an AI model designed for prostate micro-ultrasound (MicroUS), that improved performance on three image-analysis tasks, especially when trained with fewer annotated scans. The researchers pretrained it on 1,736,391 unlabeled images from 780 patients, then evaluated it on prostate capsule segmentation, urethra segmentation and cancer-lesion localization.
The work, posted on medRxiv, is a preprint. MicroUS is a high-resolution imaging method used in prostate cancer diagnosis and real-time biopsy guidance. Automated analysis could help identify structures and suspicious regions in these images, but building such systems can require experts to label large numbers of examples.
How the model was trained and evaluated
The researchers used a DINOv2-style self-supervised approach to pretrain muFM. In this kind of training, a model learns patterns from images that do not have task-specific labels; it can then be fine-tuned for a particular task using labelled examples.
The team evaluated the model in three cohorts: a public cohort of 75 patients for prostate capsule segmentation, an institutional cohort of 105 patients for urethra segmentation, and an institutional cohort of 190 patients for lesion localization. They compared muFM with randomly initialized models and foundation models pretrained on natural, medical and ultrasound images.
For segmentation, they measured Dice scores, which indicate overlap between a model's outline and the reference annotation, and the 95th-percentile Hausdorff distance (HD95), a measure of boundary disagreement. Lesion localization was assessed using free-response receiver operating characteristic analysis, which relates sensitivity to false-positive detections.
Results with limited annotations
When fine-tuned using 10% of the training scans, muFM achieved a Dice score of 93.8% for prostate capsule segmentation, compared with 87.9% for the strongest comparator. For urethra segmentation, its score was 58.4%, compared with 47.0%. The corresponding HD95 values for muFM were 2.0 mm and 4.3 mm, respectively.
For lesion localization, the authors report that muFM reached approximately 0.75 sensitivity at two false-positive detections per 3D volume and outperformed the other models. Across all three tasks, the reported pattern was strongest when annotation was limited. This suggests that pretraining on images from the target imaging modality may help reduce the amount of task-specific labelling needed to develop image-analysis models.
The study measured image-analysis performance: segmentation and lesion-localization metrics, rather than patient-level diagnostic or care outcomes. Its authors identify retrospective, single-institution pretraining and the lack of explicit 3D modelling as limitations. Whether performance carries over to other institutions and clinical workflows remains to be established.
The medRxiv posting is a preprint. It also discloses that some authors had financial, consulting, funding or employment ties to Siemens Healthineers or Exact Imaging; the latter manufactures the ExactVu MicroUS system used in the study.