Data-Efficient Prostate Micro-Ultrasound Image Analysis using a Modality-specific Foundation Model <sup>⋆</sup>.
Authors
Abstract
Micro-ultrasound (MicroUS) is an emerging high-resolution imaging modality for prostate cancer diagnosis and real-time biopsy guidance, but its automated analysis is limited by labor-intensive expert annotation. We aimed to develop and evaluate a foundation model for data-efficient MicroUS image analysis. We developed <i>µ</i> FM, a MicroUS-specific foundation model pretrained using DINOv2-style self-supervised learning on 1,736,391 unlabeled images from 780 patients. The model was fine-tuned and evaluated in three downstream cohorts for prostate capsule segmentation (public, 75 patients), urethra segmentation (institutional, 105 patients), and prostate cancer lesion localization (institutional, 190 patients). Under limited annotation, <i>µ</i> FM was compared with randomly initialized models and foundation models pretrained on natural, medical, and ultrasound images. Evaluation used the Dice scores and 95th-percentile Hausdorff distance (HD95) for segmentation and free-response receiver operating characteristic analysis for lesion localization. <i>µ</i> FM improved performance across all three tasks, with the largest gains under limited annotation. With 10% of training scans, <i>µ</i> FM achieved Dice scores of 93.8% for prostate capsule segmentation and 58.4% for urethra segmentation, versus 87.9% and 47.0% for the strongest comparators. Corresponding HD95 values were 2.0 and 4.3 mm. For lesion localization, <i>µ</i> FM outperformed other models with approximately 0.75 sensitivity at two false-positive detections per 3D volume. Limitations include retrospective single-institution pretraining and lack of explicit 3D modeling. MicroUS-specific self-supervised pretraining improved data-efficient image analysis across prostate and urethra segmentation, and lesion localization. This approach may reduce annotation requirements for developing automated tools in MicroUS image analysis. <b>What does the study add?:</b> This study evaluates a deep learning model pretrained on a large-scale dataset of unlabeled prostate MicroUS images. The pretrained model was adapted to three clinically relevant image-analysis tasks: prostate capsule segmentation, urethra segmentation, and prostate cancer lesion localization. Substantial performance gains over other models were observed despite limited training data, suggesting that MicroUS-specific self-supervised pretraining may reduce expert annotation requirements for developing automated image-analysis tools.<b>Patient Summary:</b> We developed a foundation model that learns from unlabeled prostate MicroUS images. It performed better with limited labeled data but requires independent validation before clinical use.