Prostate cancer is one of the most common cancers in men. Diagnosis relies on multiparametric MRI (mpMRI) to identify suspicious regions, followed by targeted biopsy. However, MRI is expensive, not universally available, and must be performed before the biopsy procedure.
Multiparametric ultrasound (mpUS), which combines standard B-mode imaging with contrast-enhanced ultrasound (CEUS), offers a real-time, lower-cost alternative that can be used during the biopsy procedure itself. When used alongside MRI-ultrasound fusion, mpUS can guide the needle to precisely the right location in the prostate.
Accurate segmentation of the prostate boundary and its internal zones is a critical step in this workflow. The prostate is divided into two main zones: the peripheral zone, where most cancers originate, and the transition zone, which surrounds the urethra. Knowing where each zone is helps target biopsies to the highest-risk areas.
Manual segmentation of the prostate from 3D ultrasound is time-consuming and subject to variability between operators. An automated algorithm that reliably identifies the prostate boundary and its zones could make ultrasound-guided biopsy more consistent, faster, and applicable to computer-aided diagnosis systems.
The algorithm is based on a U-Net convolutional neural network, one of the most widely used architectures for medical image segmentation. Rather than processing the full 3D ultrasound volume directly, the system converts each 3D image into a set of 2D radial planes sliced at 10-degree intervals around the central axis, yielding 18 cross-sectional images per volume.
The U-Net processes each 2D plane and predicts prostate boundaries, then the individual plane results are reconstructed back into a 3D segmentation. This approach reduces computational complexity while preserving the 3D structure of the organ. The network contains approximately 47 million trainable parameters.
Two separate neural networks were trained: one optimized for CEUS images and one for B-mode images. CEUS imaging involves injecting a contrast agent to highlight blood flow patterns; B-mode is standard grayscale ultrasound. Training separate networks acknowledges that these two modalities show tissue with very different visual characteristics.
The training dataset consisted of 259 three-dimensional mpUS images from men with clinical suspicion for prostate cancer, collected in a prospective multicenter study (NCT04605276). Manual segmentations drawn by experienced radiologists served as the ground truth for training and evaluation.
For whole-prostate segmentation, the algorithm achieved a Dice Similarity Coefficient (DSC) of 0.91 on CEUS images and 0.94 on B-mode images in 3D. The DSC ranges from 0 (no overlap) to 1 (perfect overlap), and values above 0.85 are generally considered clinically acceptable for prostate segmentation.
The mean surface distance, which measures how far the algorithm's boundary is from the manually drawn boundary on average, was less than 1.5 mm for CEUS and less than 1 mm for B-mode. These distances are small relative to the prostate's size, indicating the algorithm tracks the true boundary closely.
A key clinical validation was comparing prostate volumes calculated from the automated segmentations against MRI-based volume measurements, which are considered more reliable. The volume correlation was R2 = 0.96, meaning automated ultrasound volumes accounted for 96 percent of the variation seen in MRI volumes, confirming strong agreement between modalities.
Processing time for a complete 3D segmentation was approximately 8 seconds, compared to the several minutes required for manual delineation. This speed makes real-time use during a biopsy procedure feasible, which is a significant practical advantage.
Segmenting the peripheral zone and transition zone within the prostate is more difficult than delineating the outer boundary, because the internal border between zones is not clearly visible on ultrasound. The algorithm achieved DSC values of 0.83 for CEUS and 0.86 for B-mode for zonal segmentation, somewhat lower than whole-prostate performance but still within an acceptable clinical range.
Qualitative assessment was performed by expert radiologists who rated each segmentation on a scale from 1 to 5. Automated segmentations received a median score of 4 out of 5 with an interquartile range of 4 to 5, compared to a median of 5 out of 5 for manual segmentations. This difference was not statistically significant (p = 0.10), meaning expert reviewers found the automated contours nearly as good as those drawn by hand.
PSA density is an important clinical measure calculated by dividing the PSA blood level by the prostate volume, used to refine cancer risk assessment. PSA density calculated from automated CEUS-based segmentations (0.18 ng/ml/cm3) closely matched values calculated from MRI (0.17 ng/ml/cm3), confirming that the algorithm produces clinically meaningful volume estimates.
B-mode consistently outperformed CEUS on quantitative metrics. A key reason is acquisition time: CEUS requires longer imaging after contrast injection, during which patients may move, introducing motion artifacts that degrade image quality. B-mode images are acquired more quickly, resulting in sharper tissue boundaries and more accurate segmentation.
Previous automated segmentation methods for prostate ultrasound have reported DSC values ranging widely from 0.75 to 0.96 depending on the dataset, imaging protocol, and architecture used. The results of this study are at the higher end of this range, particularly for B-mode, placing the algorithm among the best-performing published methods for 3D transrectal ultrasound segmentation.
The most challenging cases were images with poor contrast between the prostate and surrounding tissue, or scans from patients with benign prostatic hyperplasia, where the enlarged transition zone can blur the boundary. Failure cases in the study were concentrated in these anatomically challenging situations.
A limitation of the study is that the algorithm was trained and validated on data from a single imaging protocol and a specific ultrasound system. Performance on images from different machines, different probe frequencies, or centers with different imaging conventions has not been tested. Prospective external validation would be needed before wide deployment.
The radial plane approach introduces a trade-off: it simplifies 3D processing but assumes the prostate is roughly centered on the imaging axis. For cases where the prostate is significantly off-center or asymmetric, this representation may miss fine details at the periphery. Future work could explore fully volumetric 3D network architectures to address this limitation.
This algorithm was developed as part of a computer-aided diagnosis (CAD) system for prostate cancer detection using mpUS. In this clinical pipeline, the segmentation step is a prerequisite: biopsy targeting, registration with MRI, and cancer probability mapping all depend on knowing the exact boundaries of the prostate and its zones.
Automating segmentation removes a bottleneck in this workflow. Currently, operators must manually delineate the prostate before a biopsy begins, adding time and introducing variability depending on the operator's experience. An algorithm that produces a consistent, accurate boundary in 8 seconds enables a smoother and more reproducible procedure.
The clinical calculation most directly supported by accurate segmentation is PSA density. Because PSA density helps distinguish between benign prostatic enlargement and cancer, accurate volume estimation directly influences clinical decision-making about whether to proceed with biopsy and how urgently to act.
In centers without MRI access, this workflow could support mpUS-guided biopsy as a standalone approach. The combination of automated segmentation, CEUS imaging, and CAD-based cancer probability maps represents a more accessible alternative to the MRI-ultrasound fusion pathway that requires expensive MRI infrastructure.
This study demonstrates that a U-Net-based deep learning algorithm can accurately and rapidly segment the prostate and its internal zones from 3D multiparametric ultrasound images, achieving DSC values of up to 0.94 for whole-prostate and 0.86 for zonal segmentation.
The strong correlation between automated ultrasound volumes and MRI measurements, combined with expert ratings that are statistically indistinguishable from manual contours, validates the algorithm's readiness for integration into clinical CAD workflows.
The 8-second processing time enables real-time use during biopsy procedures, overcoming the primary practical limitation of manual segmentation. This speed advantage, combined with the consistency of algorithmic contouring, could reduce operator dependence and standardize biopsy guidance across centers.
Future development should focus on external validation in multi-center settings with different ultrasound hardware, improvement of zonal boundary detection, and integration of the segmentation output into the downstream cancer detection components of the CAD system to demonstrate end-to-end clinical benefit.