Automated Prostate Gland Segmentation in MRI Using nnU-Net

Abstract: Accurate segmentation of the prostate gland in multiparametric MRI (mpMRI) is a fundamental step for a wide range of clinical and research applications, including image registration, volume estimation, and radiomic analysis. However, manual delineation is time-consuming and subject to inter-observer variability, while general-purpose segmentation tools often fail to provide sufficient accuracy for prostate-specific tasks. In this work, we propose a dedicated deep learning-based approach for automatic prostate gland segmentation using the nnU-Net v2 framework. The model leverages multimodal mpMRI data, including T2-weighted imaging, diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps, to exploit complementary tissue information. Training was performed on 981 cases from the PI-CAI dataset using whole-gland annotations, and model performance was assessed through 5-fold cross-validation and external validation on an independent cohort of 54 patients from Hospital La Fe. The proposed model achieved a mean Dice score of 0.96 +/- 0.00 in cross-validation and 0.82 on the external test set, demonstrating strong generalization despite domain shift. In comparison, a general-purpose approach (TotalSegmentator) showed substantially lower performance, with a Dice score of 0.15, primarily due to under-segmentation of the gland. These results highlight the importance of task-specific, multimodal segmentation strategies and demonstrate the potential of the proposed approach for reliable integration into clinical research workflows. To facilitate reproducibility and deployment, the model has been fully containerized and is available as a ready-to-use inference tool.

• The paper introduces a specialized nnU-Net model that integrates T2, DWI, and ADC sequences to achieve near-perfect prostate gland segmentation.
• It leverages a 3D full-resolution configuration with a ResNet encoder and is validated through balanced sensitivity and precision across cross-validation and external cohorts.
• The approach outperforms general-purpose segmentation methods, offering robust and reproducible results crucial for clinical and radiomics applications.

Automated Prostate Gland Segmentation in MRI Using nnU-Net: An Expert Summary

Problem Statement and Context

Accurate segmentation of the prostate gland in multiparametric MRI (mpMRI) underpins numerous clinical and research activities, including registration, volumetric analysis, and radiomics, yet manual delineation is hampered by labor intensity and inter-observer variability. Generic segmentation methods, while broadly applicable across anatomical structures, fail to meet the accuracy requirements for prostate-specific segmentation, primarily due to their single-modality design and the challenging morphological variability of the gland. The paper addresses these limitations by designing a specialized, multimodal deep learning pipeline for robust, automated prostate gland segmentation.

Methodology

The authors implemented a task-tailored nnU-Net v2-based segmentation model capitalizing on 3D full-resolution configuration with a ResNet encoder, enabling the preservation of volumetric context and high-frequency anatomical detail. The approach integrates T2-weighted, DWI, and ADC sequences to capture complementary tissue characteristics, outperforming single-sequence pipelines.

The model was trained from scratch using 981 mpMRI volumes from the PI-CAI dataset, annotated with Bosma22b whole-gland expert masks. Cross-validation (5-fold) with balanced Dice and Cross-Entropy loss guided optimization. For external evaluation, 54 cases from a heterogeneous Hospital La Fe cohort, carefully preprocessed to correct for variable resolution, denoising, intensity inhomogeneity, and multimodal misalignment, provided a rigorous testbed.

Segmentation quality was quantified by Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Hausdorff Distance (HD), sensitivity, and precision.

Results

The proposed model yielded a mean Dice score of 0.96 ± 0.00 and an IoU of 0.93 ± 0.01 in cross-validation within the PI-CAI dataset, indicating near-perfect spatial congruence with expert ground truth. The Hausdorff Distance remained low (4.60 ± 0.75 mm), signifying sharp and accurate boundary delineation, with balanced sensitivity and precision (0.96 ± 0.01), demonstrating comprehensive recovery of gland structure without over-segmentation.

On the external La Fe cohort, the model maintained strong generalization (Dice 0.82, IoU 0.70, HD 11.35 mm, sensitivity 0.87, precision 0.78), despite pronounced dataset domain shift across imaging vendors and acquisition protocols. In direct comparison, TotalSegmentator—a state-of-the-art general-purpose method—achieved a Dice of only 0.15 and sensitivity of 0.10, with frequent severe under-segmentation. Although its precision was high (0.91), this merely reflected its tendency to segment a very small (and mostly correct, but incomplete) portion of the prostate.

Claims and Implications

The study makes several noteworthy claims, supported by strong quantitative evidence:

• Task-specific, multimodal segmentation substantially outperforms general-purpose segmentation frameworks in prostate gland delineation.
• Multimodal input (T2, DWI, ADC) is crucial for robust performance under varying acquisition protocols and clinical conditions.
• Balanced sensitivity and precision are essential for clinically valid segmentation, whereas generic models may over-optimize for one at the expense of the other, leading to under-segmentation.

For practical deployment, the authors have containerized their model to enable reproducible research and scalable inference, facilitating direct application in clinical and translational radiomics pipelines.

Theoretical and Practical Implications

This work demonstrates that the nnU-Net framework, when properly configured with multimodal MRI data and expert annotations, yields reliable gland segmentation that is both reproducible and generalizes to external cohorts. The evident domain robustness and high segmentation accuracy are critical for quantitative mpMRI workflows and radiomics, supporting downstream tasks such as automated cancer detection, active surveillance, and treatment planning.

From a methodological standpoint, the results emphasize the nontrivial gap between general-purpose and task-optimized segmentation, confirming the necessity for customized architectures and modality fusion in challenging anatomical contexts. The findings also advocate for the systematic inclusion of external, heterogeneous validation cohorts in medical imaging AI studies.

Future Directions

Anticipated future developments include integration with domain adaptation techniques to further mitigate residual cross-site generalization loss, and expansion towards lesion-level segmentation and cross-modality learning. The availability of a robust, reproducible segmentation tool may accelerate the adoption of quantitative imaging biomarkers and support multi-institutional radiomics studies requiring harmonized gland masks.

Conclusion

This study establishes that nnU-Net-based, multimodal architectures deliver high-accuracy, reproducible prostate gland segmentation, significantly surpassing general-purpose alternatives, particularly in real-world clinical settings characterized by imaging heterogeneity. The presented pipeline sets a technical foundation for robust, automated prostate MRI workflows, with immediate translational relevance for AI-aided prostate cancer research and clinical practice.