Automated Prostate Gland Segmentation in MRI Using nnU-Net

1. Problem Statement

Accurate segmentation of the prostate gland in multiparametric MRI (mpMRI) is critical for clinical workflows, including image registration, volume estimation, radiomic feature extraction, and region-of-interest definition. However, current practices face two main challenges:

• Manual Delineation: It is time-consuming, labor-intensive, and suffers from significant inter-observer variability, limiting reproducibility and scalability.
• General-Purpose Tools: Existing automated solutions, such as TotalSegmentator, are designed for broad multi-organ segmentation across various imaging modalities. The authors argue these tools lack the specificity required for the prostate gland, often failing to utilize the complementary information provided by mpMRI sequences (T2W, DWI, ADC) and resulting in poor accuracy (specifically under-segmentation) for this specific anatomical structure.

2. Methodology

Data Sources

• Training Dataset: The model was trained on 981 cases from the PI-CAI (Prostate Imaging – Cancer AI) challenge dataset. This dataset includes mpMRI studies with T2-weighted (T2W), high b-value Diffusion-Weighted Imaging (DWI ≥1000 s/mm²), and Apparent Diffusion Coefficient (ADC) maps. Annotations were provided as whole-gland masks (Bosma22b).
• External Validation Dataset: An independent cohort of 54 patients from Hospital La Fe (Spain) was used for testing. This dataset was collected retrospectively (2015–2023) across multiple vendors (GE, Siemens, Philips) and field strengths (1.5T and 3T). Crucially, these cases included expert-validated ground truth segmentations.

Preprocessing

• PI-CAI (Training): Minimal preprocessing was applied. Images were spatially resampled to a common reference grid using the T2W image as the anatomical reference to ensure voxel-wise correspondence.
• Hospital La Fe (Testing): A rigorous harmonization pipeline was employed to handle heterogeneous acquisitions:
  • Denoising: Anisotropic diffusion filtering for T2W (SimpleITK) and SUSAN filtering for DWI (Nipype/FSL).
  • Bias Correction: N4 bias field correction for intensity inhomogeneity.
  • Registration: Rigid registration within DWI volumes and affine registration between T2W and DWI sequences using Elastix.
  • ADC Computation: ADC maps were generated from co-registered diffusion data.

Model Architecture

• Framework: The authors utilized nnU-Net v2, a self-configuring deep learning framework.
• Configuration: A 3D full-resolution mode with a ResNet-based residual encoder backbone was selected to exploit volumetric context while preserving high spatial detail.
• Input: The model accepts multimodal inputs (T2W, DWI, and ADC) to leverage synergistic tissue information.
• Training Strategy:
  • Trained from scratch on the 981 PI-CAI cases.
  • Loss Function: A composite loss combining Dice Loss and Cross-Entropy (equal weighting).
  • Validation: 5-fold cross-validation was used during training.
  • Duration: 250 epochs.

Evaluation Metrics

Performance was assessed using:

• Dice Similarity Coefficient (DSC)
• Intersection over Union (IoU)
• Hausdorff Distance (HD)
• Sensitivity and Precision

3. Key Contributions

1. Task-Specific Multimodal Approach: Unlike general-purpose tools, this work demonstrates the necessity of training a dedicated model that explicitly fuses T2W, DWI, and ADC sequences for prostate segmentation.
2. Rigorous External Validation: The study validates the model on a real-world, multi-vendor, multi-field-strength clinical cohort (Hospital La Fe), providing a realistic assessment of generalization capabilities beyond the training distribution.
3. Benchmarking Against General Tools: The paper provides a direct, quantitative, and qualitative comparison against TotalSegmentator, highlighting the limitations of "one-size-fits-all" models for specific clinical tasks.
4. Reproducibility and Deployment: The model is fully containerized (Docker), providing an end-to-end automated inference tool available via a public repository, facilitating immediate integration into clinical research workflows.

4. Results

Internal Performance (PI-CAI Cross-Validation)

The model demonstrated exceptional performance on the training distribution:

• Mean Dice Score: 0.96 ± 0.00
• Mean IoU: 0.93 ± 0.01
• Hausdorff Distance: 4.60 ± 0.75 mm
• Sensitivity/Precision: Both 0.96 ± 0.01

External Performance (Hospital La Fe Cohort)

Despite domain shift (different scanners, protocols, and patient demographics), the model maintained strong generalization:

• Proposed Model:
  • Dice: 0.82
  • IoU: 0.70
  • Hausdorff Distance: 11.35 mm
  • Sensitivity: 0.87 | Precision: 0.78
• TotalSegmentator (Baseline):
  • Dice: 0.15 (Significantly lower)
  • IoU: 0.10
  • Hausdorff Distance: 23.83 mm
  • Sensitivity: 0.10 | Precision: 0.91

Analysis of Comparison:
While TotalSegmentator showed high precision (0.91), its extremely low sensitivity (0.10) indicated severe under-segmentation, failing to capture the full extent of the prostate gland. In contrast, the proposed model achieved a balanced trade-off, successfully delineating the gland's boundaries with high completeness.

5. Significance

This work underscores that task-specific, multimodal deep learning strategies are superior to general-purpose segmentation tools for complex clinical applications like prostate cancer analysis.

• Clinical Impact: The proposed tool offers a reliable, automated solution for prostate gland delineation, reducing the burden of manual annotation and improving the consistency of downstream tasks (radiomics, volume estimation).
• Domain Shift Resilience: The model's ability to maintain a Dice score of 0.82 on an external, heterogeneous clinical cohort suggests it is robust enough for real-world deployment.
• Open Science: By releasing the containerized model and detailed code, the authors enable other researchers to reproduce results and integrate the tool into their own pipelines without technical barriers.