Prostate volume (PV) is a critical measurement in prostate cancer care. Combined with PSA levels, it enables calculation of PSA density -- a key risk stratification tool that helps clinicians decide whether a biopsy is warranted. Accurate PV assessment begins with reliable prostate segmentation on MRI.
The PI-RADS 2.1 guidelines recommend acquiring T2-weighted MRI in the axial plane plus at least one orthogonal plane (sagittal or coronal). However, real-world imaging protocols vary substantially across institutions. Scan time constraints, equipment differences, and clinical priorities lead many centers to use abbreviated protocols that acquire only the axial plane, omitting the orthogonal views that standard guidelines recommend.
Most deep learning models for prostate segmentation are trained on datasets with complete multi-planar acquisitions and assume those same planes will be available at inference time. When deployed at institutions using axial-only abbreviated protocols, these models often degrade in performance -- creating a critical gap between research benchmarks and real-world utility.
This study addresses that gap by developing a model that can accurately estimate prostate volume regardless of how many imaging planes are available -- using a novel knowledge-based (KB) contrastive loss that embeds radiological practice directly into the model's training objective.
This retrospective multicenter study used data from two sources. The publicly available PI-CAI (Prostate Imaging: Cancer AI) dataset contributed 362 patients for the final evaluation cohort (from an initial pool of 1,500 exams); an additional 204 patients from the PI-CAI ProstateX subset were used exclusively for model development and excluded from evaluation. A separate local institutional dataset contributed 46 patients for external validation.
The internal evaluation cohort had a mean age of 66.5 years and mean prostate volume of 59.3 ml, representing a wide range of prostate sizes with 55% having volumes above 50 ml. The external cohort consisted entirely of larger prostates (mean 95.7 ml), providing a more challenging external validation set. Biopsy-confirmed prostate cancer diagnoses were required for all included patients.
Axial whole-gland prostate segmentations served as ground truth for model training. In the PI-CAI dataset, some segmentations came from expert radiologists and others from a previously validated DL algorithm under radiologist supervision. For the local dataset, a radiology resident performed manual annotations using ITK-SNAP software. Reference prostate volumes were calculated using the ellipsoid formula as defined in PI-RADS 2.1, serving as the clinical standard for comparison.
The model is built on a customized 2D nnU-Net framework with three parallel encoder branches, each processing a different T2-weighted plane: axial, sagittal, and coronal. All encoders share weights. A key design choice is that only the axial encoder connects to the decoder through skip connections -- meaning the final segmentation is driven entirely by axial information, just as in standard abbreviated-protocol workflows.
The sagittal and coronal encoders do not directly influence the segmentation output but do contribute to a shared latent representation space during training. Their purpose is to enrich the model's understanding of prostate anatomy by providing complementary spatial context -- mirroring how radiologists use orthogonal views to refine their interpretation of axial slices, even though the axial plane is the primary reference.
The knowledge-based (KB) contrastive loss is the central innovation. During training, it applies a cosine similarity-based alignment penalty between the encoder outputs of corresponding axial, sagittal, and coronal slices. The model is explicitly trained to recognize that these different views all represent the same underlying anatomy. This consistency constraint allows the model to learn richer representations from orthogonal views during training, enabling it to segment accurately from axial data alone at inference. The KB loss combines this contrastive term (weighted at 0.25) with the standard nnU-Net dice and cross-entropy losses.
Five independently trained models were created via fivefold cross-validation, and their predictions were averaged at inference to generate final segmentations. All training was conducted on a single NVIDIA A100 GPU for 300 epochs using TensorFlow, making the computational requirements practical for research institutions.
With full multi-planar input (axial + sagittal + coronal), the KB model achieved a Dice Similarity Coefficient of 88.5% and Relative Volume Difference (RVD) of -8.0% in the internal cohort -- significantly better than the fully supervised reference model trained on complete multi-planar data (80.7% DSC). The Average Surface Distance of 0.54 mm indicates highly accurate boundary delineation.
Under axial-only abbreviated protocol conditions, the KB model maintained a DSC of 78.0% and RVD of -11.9% -- matching the performance of the fully supervised reference model despite using only one imaging plane. Statistical tests showed no significant difference between the two models under axial-only conditions (p=0.569 for RVD, p=0.683 for volume). This confirms the core hypothesis: the KB model performs as well as a model that was trained and tested with complete multi-planar data, even when only axial data is available.
External validation on the 46-patient cohort (with larger, more challenging prostates) confirmed and extended these findings. The KB multi-planar model achieved a DSC of 91.6% -- substantially better than the reference model at 77.3%. In the axial-only condition, the KB model still matched the reference model (78.8% vs 77.3%), demonstrating robust generalization across scanner types and institutions.
Agreement between KB-model-estimated volumes (PVKB) and the PI-RADS 2.1 ellipsoid formula reference (PVref) was assessed using the Intraclass Correlation Coefficient (ICC), Bland-Altman analysis, and scatter plots. ICC values above 0.90 are considered excellent, and the KB model achieved this in the internal dataset under both multi-planar (ICC 0.98) and axial-only (ICC 0.98) conditions.
In the external cohort, ICC values were 0.90 for multi-planar and 0.88 for axial-only inputs -- rated as excellent and good respectively. Bland-Altman plots showed that the KB model's mean difference and 95% limits of agreement were narrower than the reference DL model under both testing conditions, meaning the KB model introduces less systematic bias and less variability when estimating prostate volume.
No evidence of systematic over- or underestimation was observed across conditions. The axial-only KB model showed slightly wider confidence bounds than the multi-planar version, as expected, but the difference was clinically acceptable. These findings confirm that PSA density calculations based on KB model volume estimates would be reliable even when institutions use abbreviated axial-only protocols -- directly supporting clinical decision-making in prostate cancer management.
The KB-loss approach represents a practical path toward sustainable clinical AI deployment. Unlike multi-planar models that require extensive annotation of orthogonal views and can only be used at fully-equipped imaging centers, the KB model requires only axial segmentations during training and adapts automatically to whatever planes are available at inference -- without retraining.
This design directly aligns with two convergent trends in clinical practice: the movement toward abbreviated MRI protocols to reduce scan time and contrast exposure, and the growing demand for AI tools that work reliably across heterogeneous institutional workflows. The study authors argue that embedding radiological reasoning into the model's learning objective -- rather than simply requiring more data -- may be a more principled and scalable approach to clinical robustness.
Limitations include the retrospective design, reliance on a single DL architecture, and the need for broader testing across more scanner types and acquisition protocols. Prospective validation and testing in fully abbreviated biparametric protocol settings would strengthen the case for deployment. Nevertheless, the consistent performance across two independent institutional datasets establishes a strong foundation for clinical translation of this approach to AI-assisted prostate volume estimation.