Accurately outlining the boundaries of the prostate on MRI images, a process called segmentation, is a critical step in prostate cancer diagnosis and treatment. It enables accurate volume measurements, guides where to direct biopsy needles, helps plan radiation therapy, and tracks treatment response over time. Without a clear delineation of the prostate, none of these downstream tasks can be performed reliably.
Manual segmentation by a radiologist is the current standard, but it is slow, time-consuming, and subject to significant inter-observer variability: different radiologists often draw slightly different boundaries around the same prostate. This inconsistency can lead to measurement errors that affect clinical decisions. Automating segmentation would make the process faster, cheaper, and more consistent.
Automatic segmentation of the prostate is technically challenging because MRI images of the prostate have inhomogeneous intensity distributions: the same tissue type can appear different shades of gray in different parts of the image or on different scanners, and the prostate boundary is often fuzzy and poorly defined, particularly at the apex and base of the gland. These properties confuse algorithms that rely purely on image intensity differences.
Deep learning, and specifically convolutional neural networks (CNNs), have shown impressive performance for medical image segmentation tasks. However, most existing approaches rely only on image appearance features and do not incorporate prior knowledge about prostate anatomy, leaving room for improvement in difficult cases where image contrast alone is insufficient.
This paper proposes an automatic prostate segmentation method that incorporates anatomical knowledge directly into the learning process. The approach builds on the well-established 3D U-Net architecture, a deep learning model designed specifically for volumetric medical image segmentation, and augments it with additional constraints derived from the known anatomy of the prostate.
The 3D U-Net processes the full three-dimensional MRI volume rather than individual 2D slices. It uses an encoder path to progressively extract high-level feature representations, and a decoder path to expand those representations back to the original image resolution, producing a voxel-by-voxel segmentation mask. Skip connections between the encoder and decoder preserve fine-grained spatial details at multiple scales, which is important for accurately capturing the prostate boundary.
Two key anatomical facts about the prostate are encoded as additional loss functions. First, location knowledge: the prostate is anatomically positioned between the rectum and the pelvic bone, placing it roughly in the center of pelvic MRI scans. A location loss penalizes the model when predicted segmentations fall outside the expected anatomical region, preventing gross errors in which the algorithm incorrectly labels distant structures as the prostate.
Second, shape prior knowledge: the prostate has a roughly walnut-shaped, elliptical cross-section. An autoencoder, which is a neural network trained to compress and then reconstruct data, is used to learn the characteristic shape of the prostate. A shape loss penalizes the model when its predicted segmentation deviates substantially from this expected shape, encouraging smooth, anatomically realistic boundaries even when the MRI image is noisy or the boundary is indistinct.
Training a deep learning model requires defining a loss function, which measures how wrong the model's predictions are. The model learns by iteratively adjusting its parameters to reduce this loss. This paper uses three loss functions combined in a weighted sum, each targeting a different aspect of segmentation quality.
The Dice loss directly optimizes the overlap between the predicted segmentation and the ground truth annotation, using the Dice similarity coefficient. Because the prostate occupies a small fraction of the total MRI volume, there is a severe class imbalance between foreground (prostate) and background voxels. Dice loss handles this imbalance better than standard cross-entropy loss, which would be dominated by the vast number of background voxels.
The location loss computes Dice loss specifically within the minimum bounding rectangle (MBR) of the expected prostate region, focusing the model's attention on the anatomically plausible location. This constraint helps the model converge faster during training and prevents it from proposing segmentations in anatomically impossible locations such as the bladder neck or rectal wall.
The shape prior loss uses KL divergence to measure the difference between the shape features of the predicted segmentation and those of the ground truth, as encoded by an autoencoder network. By penalizing unrealistic shapes, this loss function encourages the model to produce smooth, convex prostate outlines rather than jagged or irregular predictions that could occur when image boundaries are ambiguous.
The method was evaluated on the PROMISE2012 challenge dataset, a public benchmark for prostate segmentation consisting of 50 MRI volumes acquired at multiple hospitals using different MRI scanners and protocols. This multi-site diversity tests whether an algorithm generalizes beyond the specific settings it was developed in, which is critical for clinical deployment.
All volumes were preprocessed to a uniform spatial resolution of 1.0 x 1.0 x 1.5 millimeters to ensure consistency. The network was trained with a fixed input volume size of 64 x 128 x 128 voxels and a batch size of 2, running for 100 training epochs with the Adam optimizer. The initial learning rate was set to 0.002 with scheduled decay.
A leave-one-out cross-validation strategy was used: for each of the 50 patients, the model was trained on the remaining 49 patients and tested on the held-out patient. This approach maximally uses the available data and provides an unbiased estimate of generalization performance, though the small dataset size remains a limitation.
Performance was measured using two standard metrics: the Dice similarity coefficient (DSC), which measures volumetric overlap between predicted and manual segmentations (1 = perfect, 0 = no overlap), and the Hausdorff distance (HD), which measures the maximum surface distance between the predicted and ground truth boundaries in millimeters, with smaller values indicating better boundary accuracy.
The proposed anatomy-guided method achieved a mean Dice similarity coefficient of 91.6% with a standard deviation of 3.26% on the PROMISE2012 dataset. This indicates very high volumetric overlap with expert manual segmentations, meaning the automated method identifies nearly the same prostate region as an experienced radiologist would.
The mean Hausdorff distance was 5.18 mm with a standard deviation of 1.61 mm. This measure of boundary accuracy indicates that the most extreme discrepancy between the automated and manual segmentation boundaries was approximately 5 mm on average, which is clinically acceptable for most applications including treatment planning and volume measurement.
To demonstrate the benefit of incorporating anatomical knowledge, the authors compared segmentation results with and without the anatomy-based loss functions. Qualitative comparison showed that predictions made with anatomy constraints produced more complete prostate regions with smoother, more anatomically realistic boundaries, while predictions without anatomy constraints were more fragmented and included more outlier predictions in anatomically implausible locations.
When compared against seven previously published segmentation methods on the same PROMISE2012 dataset, the proposed anatomy-guided approach achieved a higher DSC than all competing methods, demonstrating the value of integrating prior anatomical knowledge into the deep learning framework beyond pure image-intensity-based learning.
This work demonstrates that incorporating structured anatomical knowledge into a deep learning segmentation framework, rather than relying solely on image features, meaningfully improves prostate MRI segmentation performance. The combined loss function approach is elegant and practical: it does not require additional labeled data or complex network architectures, but instead leverages well-known anatomical facts about prostate location and shape.
High-quality prostate segmentation is the foundation of the entire MRI-guided prostate cancer care pathway. Automated segmentation with accuracy approaching that of expert radiologists could substantially reduce the workload on radiology departments, enable consistent measurements across follow-up scans, and make high-quality prostate MRI analysis accessible in settings without subspecialty-trained prostate imaging radiologists.
The method's demonstrated generalizability across different scanners and acquisition protocols on the PROMISE2012 multi-site dataset is encouraging for real-world deployment, where patients are scanned at different facilities with different equipment. This robustness is critical for any clinical AI tool to be adopted beyond the institution where it was developed.
Future work should validate the method on larger, more diverse patient cohorts and evaluate performance specifically in zones of the prostate where segmentation is most challenging: the apex and base, where the boundary with surrounding structures is least distinct. Extension to zonal segmentation, distinguishing the peripheral zone from the transitional zone, would further increase the clinical utility of the approach.