Multiparametric MRI (mpMRI) is now the leading imaging method for detecting, locating, and characterizing prostate cancer before biopsy. The PI-RADS scoring system uses a five-point scale to rate the probability that a suspicious MRI finding represents a clinically significant cancer -- and the score depends critically on which zone of the prostate the lesion is located in.
The prostate has two main zones: the peripheral zone (PZ), which surrounds the back and sides of the gland and is where most cancers arise, and the transition zone (TZ), the inner region that also harbors benign prostatic hyperplasia (BPH). PI-RADS specifies that different MRI sequences have different diagnostic importance depending on the zone: T2-weighted imaging is dominant for TZ lesions, while diffusion-weighted imaging takes precedence for PZ lesions. Misidentifying the zone changes the scoring rules and can lead to incorrect assessment.
Radiologists also use standardized sector maps -- diagrams dividing the prostate into up to 39 named regions -- to communicate lesion locations across clinical teams for biopsy targeting and treatment planning. Consistent and accurate zonal and sector-level localization is therefore fundamental to prostate cancer management, yet manual segmentation is time-consuming and subject to significant inter-radiologist variability.
Additionally, 3D T2-weighted MRI is beginning to replace traditional 2D sequences because it offers faster scanning, thinner image slices (allowing better anatomical detail), and easier fusion with other imaging modalities. However, most existing automated segmentation methods were developed exclusively for 2D sequences and cannot directly handle 3D data. This study aimed to address both the zonal segmentation and the 2D-to-3D gap simultaneously.
The authors developed UFNet, a two-stage deep learning framework built on the U-Net architecture -- a widely used convolutional neural network design for medical image segmentation. The first stage (global location network) rapidly identifies the approximate position and bounding box of the prostate within the full MRI image. The second stage (zonal segmentation network) then operates at higher resolution within that bounding box to precisely delineate the whole gland (WG), transition zone (TZ), and peripheral zone (PZ).
A key technical innovation was the handling of image anisotropy -- the fact that prostate MRI images have different spatial resolutions in different directions (fine in-plane, coarser through-plane). UFNet uses anisotropic convolution kernels that separate in-plane and through-plane processing, preventing the network from treating artificially stretched voxel dimensions as true 3D isotropic data.
UFNet also incorporates two types of attention mechanisms that help the network focus on the most diagnostically relevant image regions. Squeeze-and-excitation modules perform channel-level attention, re-weighting feature maps according to their informational importance. Attention gates perform spatial attention, highlighting the most relevant spatial locations within skip connections (links between network layers at different scales).
A novel set of partition loss functions was designed to mathematically enforce a biologically correct constraint: the peripheral zone and transition zone together must exactly fill the whole gland, with no gaps or overlaps. This prevents the network from generating anatomically impossible outputs where TZ and PZ overlap or leave parts of the prostate unassigned.
The primary training and evaluation dataset consisted of 131 high-resolution 3D T2-weighted MRI scans acquired at two Paris hospitals (Pitie-Salpetriere and Tenon) using two different scanner brands (Siemens and GE). Using two scanner types intentionally introduced the kind of image variability seen in real clinical practice, making the model more likely to generalize.
A key methodological strength was the generation of seven independent radiologist segmentations for each of the 40 test set cases -- contributed by radiologists of three experience levels: experts (more than 1,000 prostate MRIs interpreted), seniors (approximately 500), and juniors (fewer than 100). A consensus segmentation derived from all seven using the STAPLE algorithm served as the gold standard, and the AI model's results were placed in context by comparing it directly to each individual radiologist's performance.
The study also evaluated on the publicly available ProstateX dataset (201 cases from a different institution), enabling comparison with previously published methods. This external validation on a completely different dataset is critical for assessing whether a segmentation model will transfer to new clinical environments.
The team developed an automated algorithm to convert the zonal segmentation into a prostate sector map -- dividing the gland into anatomical regions based on longitudinal (apex/midgland/base), anteroposterior (anterior/posterior), and transverse (left/right) axes. This enabled the study to assess not just whether zones were correctly identified, but whether cancer lesion locations would be correctly communicated to biopsy and treatment planning teams.
UFNet-E (the ensemble of five networks) achieved Dice similarity coefficient (DSC) scores of 93.9% for the whole gland, 91.0% for the transition zone, and 79.1% for the peripheral zone on the private 3D dataset. Higher Dice scores indicate better overlap between the AI segmentation and the expert consensus. These results were statistically significantly better than the baseline UNet model across all zones (p less than 0.001).
On the ProstateX public dataset (2D scans), UFNet-E achieved DSC scores of 90.9% for the whole gland, 86.8% for TZ, and 78.4% for PZ -- values that are comparable to or better than competing state-of-the-art methods tested on the same dataset. The consistency across both datasets demonstrates the model's generalizability to different scanning protocols and patient populations.
When UFNet-E was ranked alongside the seven radiologists based on proximity to consensus, it placed fifth out of eight (between the five radiologists), performing significantly better than one expert radiologist on both TZ and PZ segmentation (p less than 0.05). The model was not as accurate as the two best radiologists overall, but it consistently exceeded junior and some senior radiologists.
The most challenging segmentation regions were the apex and base of the prostate -- the top and bottom slices -- where tissue heterogeneity, ambiguous boundaries, and unusual shapes make accurate delineation difficult for humans and algorithms alike. This finding is consistent across both human and automated segmentation studies in the literature.
For the test set prostates containing a suspicious lesion, UFNet's automated sector map correctly identified the zone (PZ or TZ) of every lesion with 100% accuracy. The correct sector within the zone was identified in 88% of cases (15 out of 17 lesions), with the two errors involving lesions near the boundary between apex and midgland regions.
On an independent private lesion dataset containing 46 clinically significant lesions, zonal accuracy was 91% (42 out of 46 correct zone assignments). The four errors all involved the same specific anatomy: lesions in the anterior peripheral zone at the TZ border, where the tissue intensity looks more like TZ than PZ -- a known challenge even for experienced radiologists.
Sector-level accuracy on the private lesion dataset was 74% (34 out of 46 correct sector assignments). This is comparable to the 74% agreement reported in published literature between human radiologists for the same type of sector map, suggesting the AI is performing at a level consistent with inter-radiologist agreement -- not worse, but not better either.
From a clinical standpoint, the observed sector errors would have minimal impact on patient care. In nearly every case where the algorithm misidentified the sector, it identified the correct zone and a directly adjacent sector, meaning biopsy targeting guidance would still direct the needle to the correct general area. The authors suggest that reporting the top two most probable sectors for each lesion -- rather than a single sector assignment -- would capture the true sector in almost all cases.
The study demonstrated that deep learning-based prostate zone segmentation is not only technically feasible but practically useful: it performs at the level of mid-tier radiologists, processes each scan in approximately 4-5 seconds, and reliably preserves the information clinicians need for cancer grading and biopsy planning.
As 3D T2-weighted MRI becomes the new radiological standard for prostate imaging (offering faster acquisition, thinner slices, and better multi-modality image fusion), having validated automated segmentation for this modality specifically is an important step. UFNet is the first published method to achieve state-of-the-art zonal segmentation on 3D sequences, bridging a critical gap.
The automated sector map generation is a novel contribution with direct clinical application. Connecting zone segmentation to standardized sector maps enables the AI to communicate lesion locations in the same clinical language that radiologists and urologists already use for biopsy guidance, follow-up reports, and multidisciplinary team discussions.
Future work should address the remaining challenges: improving PZ segmentation at the gland's apex, reducing errors for lesions at the TZ-PZ boundary, and adopting the updated 39-region PI-RADS 2.1 sector map. Integration with full lesion detection and PI-RADS scoring pipelines would create a complete automated diagnostic system for prostate cancer MRI review.