To develop a novel transparent and lightweight machine learning model, the Green Learning (GL), for automated prostate segmentation (PS) and clinically significant prostate cancer (csPCa) detection on magnetic resonance imaging (MRI). Men who underwent 3-T MRI and prostate biopsy (PBx) were identified. MRI was acquired and interpreted according to the Prostate Imaging-Reporting and Data System (PI-RADS), version 2 or 2.1. The GL was created to automate PS and csPCa detection on biparametric MRI. The performance was compared to the standard-of-care radiologists using PI-RADS, and a conventional deep learning (DL) U-Net model as benchmarking. The PS performance was evaluated by the Dice similarity coefficient (DSC). The area under the curve (AUC) for patient-level csPCa detection was assessed. Model size and computational workload, measured by floating point operations (FLOPs), were reported. A total of 602 MRIs were randomly divided for training (N = 483) and testing (N = 119). Overall, 224 patients had csPCa on PBx. The median DSC for PS was higher for GL than U-Net (0.91 vs 0.88, P < 0.001). The AUC for csPCa detection of GL was similar to PI-RADS (0.75 vs 0.76, P = 0.8) and U-Net (vs 0.74, P = 0.3). A combination of GL and PI-RADS showed a higher AUC of 0.81 than PI-RADS alone (P = 0.02). Compared with U-Net, the GL had smaller magnitude parameters (1.21× 106 vs 177× 106) and less computational workload (9.8× 109 vs 1027× 109 FLOPs). A novel GL model fully automatically detects csPCa on prostate biparametric MRI with comparable performance to PI-RADS and DL. Combined with PI-RADS, GL significantly improves csPCa detection.

Unsupervised domain adaptation for medical image segmentation by pre-training and 3D multi-scale feature fusion

Deep learning (DL) has driven major progress in medical imaging diagnosis. However, its effectiveness is often limited by the scarcity of large annotated datasets and the poor generalization of models to small, out-of-distribution (OOD) data. Self-supervised learning (SSL) and transfer learning offer promising solutions: SSL enables representation learning from unlabeled data via pretext tasks, while transfer learning helps adapt models to new domains with limited labeled data. In this study, we evaluate SSL-based strategies for detecting and segmenting clinically significant prostate cancer (csPCa) in MRI. We investigate the impact of key design decisions-including model architectures, pretext tasks, contrastive learning methods, and downstream tasks-using one medium-sized pre-training dataset (PI-CAI) and two small OOD target datasets: Prostate158 and ChiPCa. We propose a three-stage training pipeline that includes SSL, supervised pre-training, and a final fine-tuning on the target small labeled dataset.Results show that our proposed full three-stage training pipeline achieves the most consistent performance in both detection and segmentation on Prostate158, whose csPCa area distribution is closer to the pre-training dataset. In contrast, for ChiPCa, whose csPCa distribution differs from the pre-training dataset, the full strategy is the best for detection but suboptimal for segmentation, where partial training stages can provide better results. In general, either UNETR or UNet can serve well for detection, but UNet architecture reports consistently better results for segmentation. These findings provide practical guidance on when multi-stage SSL pipelines are most beneficial and how dataset similarity and architectural choice influence prostate segmentation performance.

To develop and validate an interpretable machine learning model integrating radiomics, deep learning (DL), and clinical features based on automated MRI segmentation for detecting prostate cancer (PCa). This retrospective multicenter study included 433 prostate patients. The internal cohort comprised 346 patients, who were randomly divided into a training set (n = 242) and an internal validation set (n = 104) at a 7:3 ratio. Automated prostate segmentation was performed on T2-weighted imaging and apparent diffusion coefficient maps using TotalSegmentator. Radiomics features were extracted and selected via mutual information, mRMR, LASSO, and Pearson correlation analysis. The DL labels were derived from a DenseNet-121 convolutional neural network. Using the eXtreme Gradient Boosting (XGBoost) algorithm, we constructed four types of prediction models: clinical models, radiomics models, DL models, and combined models integrating all three feature types (clinical, radiomics, and DL features). Model performance was evaluated using area under the curve (AUC) and related metrics. An external test set of 87 patients was used to validate the predictive performance of each model. Finally, SHapley Additive exPlanations (SHAP) were applied to enhance model interpretability and quantify the impact of individual features. The combined model integrating radiomics, DL, and clinical features achieved the highest AUC of 0.902 (95% CI 0.841-0.962) in the external test set, outperforming individual models. SHAP analysis revealed prostate-specific antigen density and DL label as dominant predictors and provided transparent global and local interpretations of model decisions. The combined model integrating clinical, radiomics, and DL features based on automated MRI segmentation, achieved high accuracy and promising clinical utility in PCa detection.

Multiparametric MRI (mpMRI) and ^68Ga-PSMA PET/CT are widely used for prostate cancer (PCa) diagnosis but remain limited by false positives and modest specificity, particularly in distinguishing benign prostate diseases (BPDs) and clinically significant PCa (csPCa). Existing studies often rely on small, single-center cohorts with limited generalizability. This study aimed to develop and externally validate a multimodal radiomics model integrating PET/CT and mpMRI for automated PCa diagnosis, and to evaluate the impact of prostate VOI delineation strategies. A total of 488 patients with suspected PCa who underwent both ^68Ga-PSMA PET/CT and mpMRI (T2 and DWI) followed by biopsy were retrospectively enrolled from two centers (366 for model development and ten-fold internal validation; 41 for external validation cohort 1; 81 for external validation cohort 2). Radiomics features were extracted from both modalities, and six classical machine learning classifiers (LR, SVM, Random Forest, Extra Trees, XGBoost, LightGBM) were trained for three tasks: (1) csPCa diagnosis, (2) overall PCa detection, and (3) comparison between expert-drawn and deep learning generated prostate VOIs. Model performance was assessed using AUC, sensitivity, specificity, accuracy, PPV, and NPV. Among 407 patients, 137 had BPD, 25 had clinically insignificant PCa, and 250 had csPCa. The multimodal PET/mpMRI radiomics model achieved the best performance with LightGBM (AUC = 0.91 internally; 0.825 externally). Automatically segmented VOIs achieved comparable diagnostic accuracy to expert annotations, with AUC differences within 3-8%. The proposed multimodal PET/CT and mpMRI ML-based model enables accurate risk stratification for prostate cancer, with strong external generalizability. Automated prostate segmentation provides comparable diagnostic performance to expert manual delineation, facilitating clinical scalability.

The annual mortality rate from prostate cancer (PCa), a common malignant neoplasm affecting middle-aged and elderly men, is on the rise. Biparametric magnetic resonance imaging (bpMRI) is indispensable to PCa imaging analysis since it can capture distinct disease-related information from two modalities that exhibit synergistic performance. The majority of state-of-the-art PCa diagnostic techniques currently available are focus on a single modality or task, neglecting the information sharing across the two modalities and task correlations inherent in multi-task learning. We provide a dual-modality image fusion and multi-task learning model that can accomplish both automatic PI-RADS grading and prostate and PCa region segmentation simultaneously. First, to extract complementary information between the prostate and PCa in bimodal images via T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) feature extraction, a shared block fusion module and an independent encoder block were developed; Subsequently, in the encoder stage, the dual visual attention module was designed to extract features from multiple receptive field and deliver more accurate contextual information, and a novel decoder was designed to effectively integrate encoder features, yielding more refined global and local detail information; Next, to capture more precise detail information during the classification task stage, a high-level feature fusion technique was developed; To address class imbalance, a multitask mixed loss function is finally suggested. The segmentation results of prostate and PCa on multiple diverse male pelvic MRI datasets demonstrate the superior performance of our proposed method. Both the basic performance evaluation and comparative model evaluation of the proposed model have validated its effectiveness in prostate and PCa segmentation as well as PI-RADS automatic grading. External validation on the independent PROMISE12 dataset further confirms the strong generalizability of our model across different institutions, scanning devices and patient cohorts.

BackgroundAccurate delineation of the prostate and intraprostatic gross tumor volume (GTV) on multiparametric MRI (mpMRI) is critical for radiation therapy planning, particularly for focal dose escalation strategies. However, interpretation of mpMRI can be challenging and prone to inter‐observer variability, especially among radiation oncologists (ROs) who may have limited training in prostate MRI interpretation. In addition, because many patients do not undergo diagnostic mpMRI before treatment, radiologist input is often absent during treatment planning, which can compromise accurate GTV delineation.PurposeTo evaluate approaches for improving focal prostate GTV delineation on mpMRI for radiotherapy planning, by assessing the impact of radiology reports on ROs contours, and comparing the performance of an automated segmentation tool (AIRC) with radiologists' contours as a potential replacement when radiologist input is unavailable.MethodsTwenty patients underwent mpMRI at our institution. Delineation of the prostate gland and GTV was performed by ROs in two phases. In the first phase, ROs segmented both structures with access only to biopsy results, indicating positive core locations. In the second phase, they repeated the segmentation with access to the radiology report. AIRC segmented both structures as well, using the same images as made available to the ROs. Lesion detection performance was evaluated using sensitivity and positive predictive value (PPV). Segmentations were compared to consensus contours from two radiologists using the STAPLE algorithm. Similarity metrics computed included Dice similarity coefficient (DSC), Hausdorff distance (HD), distance from center, and relative volume difference.ResultsProstate gland segmentation showed minimal inter‐observer variability and high agreement with the reference, with a mean DSC of 0.89±0.02 for ROs and 0.90±0.03 for AIRC. In contrast, GTV segmentation demonstrated substantial variability and a high number of false positives and missed lesions when ROs lacked access to the radiology report. PPV significantly improved during the second attempt with access to the radiology report (p=0.03), and similarity metrics improved significantly across all measures (p<0.01). Sensitivity also increased in the second attempt, although not significantly (p=0.0625). AIRC performance was comparable to the ROs' first attempt.ConclusionsHigh inter‐observer variability in GTV segmentation was observed when ROs lacked access to the radiology report, underscoring the difficulty of mpMRI interpretation. Providing radiology reports and establishing standardized guidelines for mpMRI interpretation may improve segmentation accuracy and consistency. An Automated segmentation tool could facilitate the first step in the planning workflow by assisting ROs in target definition.

Abstract No. 49 ▪ FEATURED ABSTRACT Automated Prostate Segmentation for MRI-Guided Transurethral Ultrasound Ablation: Validation and Real-World Experience

Accurate early detection of clinically significant prostate cancer is crucial for improving patient outcomes. However, traditional diagnostic methods such as Digital Rectal Exam and Prostate-Specific Antigen (PSA) tests often lack the sensitivity and specificity needed for effective diagnosis. This study presents an AI-based approach for csPCa classification using MRI data, incorporating both the PI-CAI Challenge dataset and a newly compiled, diverse BIMCV Prostate dataset comprising over 9000 MRI sessions from 16 healthcare centers in the Valencian Region. The methodology includes a robust preprocessing pipeline, featuring prostate segmentation with a custom-trained nnUNet model, and utilizes a 3D variant of EfficientNet-B7. To ensure robustness, we employed a transfer learning strategy where five models pretrained on PI-CAI were fine-tuned on the BIMCV dataset and aggregated using a stacked meta-learner. This ensemble approach yielded a Receiver Operating Characteristic Area Under the Curve of 0.816 on the independent hold-out set, significantly outperforming a non-pretrained baseline (AUC 0.71). Furthermore, we demonstrated that synthesizing missing ADC maps using a mono-exponential model serves as an effective data augmentation strategy, preventing data loss without introducing domain shift. Interpretability techniques such as occlusion sensitivity and guided backpropagation were employed to provide insights into the model's decision-making process, enhancing transparency. This research highlights the potential of AI-enhanced MRI techniques in advancing csPCa detection and diagnosis.

Dynamic volumetric MRI was used to non-invasively assess urethral biomechanics in a single healthy male subject during voiding. Volumetric lower urinary tract (LUT) images were obtained throughout the voiding effort using the uro-dynamic MRI methodology. These were subsequently segmented using MIMICS. Segmented urethral volumes were divided into prostatic, membranous, penile and bulbous urethra and were used to quantify urethral diameters over time, length and different phases of flow. Idealized instantaneous resistances were calculated for the posterior (prostatic and membranous) urethra. The analysis was separated into time and length averaged, and step-by-step analysis. Step-by-step analysis revealed more variation in the initial flow and terminal flow than during robust flow. Time and length averaged analysis showed the membranous urethra had the narrowest lumen and greatest resistance to flow. Correlation of urethral diameters with flow showed strong correlations with both the prostatic and membranous urethral segments. This single-subject study confirms the potential of uro-dynamic MRI to provide noninvasive assessment of lower urinary tract anatomy and urethral biomechanics during voiding.

ProAttNet: a novel network of prostate segmentation with multi-attention residual U-Net using magnetic resonance images

Federated learning (FL) enhances data privacy and compliance with data regulations by enabling multiple decentralized parties to collaboratively train machine learning models without sharing their data. This makes it an ideal paradigm for the healthcare domain. Despite its inherent privacy-by-design, FL faces performance and convergence challenges when dealing with non-independent and identically distributed (non-IID) data. Although previous studies have primarily addressed the challenges of skewed label distribution across clients, which are more effective for classification, we focus in this paper on the much less explored challenge of multi-domain FL in medical image segmentation tasks, where client data originate from different domains with varying feature distributions. To address this problem, we propose FedGA, which performs heuristic, gradient-free optimization on the server side after local model aggregation in order to optimize the global model by using a genetic algorithm. Empirical results on breast lesion segmentation from ultrasound images and prostate segmentation from T2w MRI images show the potential of FedGA in federated learning schemes. Specifically, FedGA improves the segmentation precision in critical boundary regions compared to existing approaches, accelerates global model convergence, reduces the total number of communication rounds required to achieve optimal performance, and offers better overall efficiency.

Prostate MRI segmentation is critical for accurate diagnosis and treatment planning but remains challenging due to the complex interplay between the peripheral zone's thin, irregular boundaries and the central gland's homogeneous textures, compounded by variability across imaging protocols. To address these challenges, we propose ProSeg, a novel deep learning framework featuring a specialized ProSeg block that integrates dual complementary processes: (1) anisotropic convolutions for precise peripheral zone boundary delineation and (2) cross-slice attention mechanisms for robust central gland texture modeling. Extensive evaluations on the Promise12 and Promise158 datasets demonstrate ProSeg's state-of-the-art performance, achieving Dice scores of 84.31% (peripheral zone) and 57.92% (central gland) on Promise12, and 83.15% (peripheral zone) and 56.38% (central gland) on Promise158, significantly outperforming existing methods. ProSeg's consistent accuracy across diverse protocols highlights its clinical potential for reliable prostate zonal segmentation in real-world settings.

This work investigates the extension of StyleGAN2-ADA to three-dimensional prostate T2-weighted (T2W) MRI generation. The architecture is adapted to operate on 3D anisotropic volumes, enabling stable training at a clinically relevant resolution of 256×256×24, where a baseline 3D-StyleGAN fails to converge. Quantitative evaluation using Fréchet Inception Distance (FID), Kernel Inception Distance (KID), and generative Precision-Recall metrics demonstrates substantial improvements over a 3D-StyleGAN baseline. Specifically, FID decreased from 114.2 to 27.3, while generative Precision increased from 0.22 to 0.82, indicating markedly improved fidelity and alignment with the real data distribution. Beyond generative metrics, the synthetic volumes were evaluated through radiomic feature analysis and downstream prostate segmentation. Synthetic data augmentation resulted in segmentation performance comparable to real-data training, supporting that volumetric generation preserves anatomically relevant structures, while multivariate radiomic analyses showed strong global feature alignment between real and synthetic volumes. These findings indicate that a 3D extension of StyleGAN2-ADA enables stable high-resolution volumetric prostate MRI synthesis while preserving anatomically coherent structure and global radiomic characteristics.

Aim: The aim of this study is to automatically segment the prostate gland, transitional zone (TZ) and periferal zone (PZ) on prostate Magnetic Resonance Imaging (MRI) using a U-net based convolutional neural network (CNN).Materials and Methods: This retrospective study included a total of 100 patients who underwent screening with a 1.5T MRI device between January and December 2020. The acquired images were evaluated by a senior radiology resident and converted to nifti format using the MedSeg.ai platform. Prostate and TZ masks were manually traced, while the remaining area (PZ) was automatically segmented by extracting the TZ mask from the prostate mask. A U-net based CNN algorithm with 7 depth layers was developed. Data from 80 patients were used for training the algorithm, with 10 randomly selected for validation. The remaining data from 20 patients were used for testing. Evaluation metrics applied on the test set included accuracy, mean and median Dice Similarity Coefficient (DSC), mean Hausdorff Distance (HSD), Mean Surface Distance (MSD), mean Relative Absolute Volume (RAV).Results: Mean DSC of 0.91 ± 0.03, 0.87 ± 0.06, 0.70 ± 0.16 and median DSC of 0.92, 0.90, 0.75 were obtained for prostate gland, TZ and PZ segmentation respectively. Mean HSD was 8.58, 9.52, 18.78, MSD was 0.92, 0.84, 1.30 and mean RAV was 3.51, 9.87, 70.57 for the segmentation of aforementioned structures.Conclusion: The developed U-net algorithm performed better in segmenting the prostate and TZ than in previous studies. While the success rate of PZ segmentation was lower, this could be attributed to various factors, as indicated by state-of-the-art methods in deep learning. This study highlights AI's promising role in automating prostate segmentation.

Background/Objectives: Active surveillance (AS) has emerged as a safe alternative to primary therapy for low- and select intermediate-risk prostate cancer (PCa), but optimal patient selection and surveillance strategies remain challenging due to limited risk stratification tools enabling early detection of lesions with high potential for histopathological progression. This study presents an integrated method for predicting prostate cancer progression within 12 months, aiming to improve AS patient selection by categorizing patients into two risk groups: rapid progressors who would benefit from immediate treatment and slow progressors suitable for AS. Methods: The risk assessment platform combines convolutional neural networks for automatic segmentation of prostate and suspicious-for-cancer lesions on multiparametric MRI (mpMRI) with logistic regression to estimate progression risk. The networks were trained on annotated lesions from radical prostatectomy specimen mapped to mpMRI. The prediction model incorporated pre-biopsy clinical variables (age, PSA, PI-RADS) and MRI-derived intratumoral radiomic features from 163 participants of a prospective clinical trial, using histopathological progression within 12 months as endpoint. Results: The clinical-radiomics model achieved an AUC of 0.84 in distinguishing rapid from slow progressors, using non-invasive monitoring techniques. In an independent test set, the model significantly improved AS patient selection, increasing negative predictive value by 18.5% compared to current standard-of-care (p < 0.001). Conclusions: The risk assessment platform shows promise for use during annual follow-up visits to reliably differentiate suitable AS candidates with stable disease from PCa patients who are likely to experience early progression.

Multi-scale attention and hypergraph neural network based deep learning framework for prostate cancer classification and segmentation

KanGMSA-Net: A novel prostate segmentation framework integrating Kolmogorov–Arnold networks and grouped multi-scale attention

The Mona Lisa 2.0 robotic platform integrates MRI/ultrasound fusion, AI-based prostate segmentation, and automated needle trajectory planning to optimize transperineal targeted biopsy (TB) precision. We report the first European experience in 10 consecutive patients undergoing robot-assisted TB, with optional systematic cores. Clinically significant prostate cancer was detected in all rTB procedures. Standard cores added limited diagnostic yield and mainly sampled perilesional "penumbra" areas. Mean biopsy duration was 12.9 min, no peri- or post-procedural complications occurred, and high-quality tissue samples were consistently obtained. These preliminary data confirm feasibility, safety, and reproducibility of Mona Lisa 2.0 robotic platform, as a new kid on the block in urologic robotic armamentarium.

High-frequency micro-ultrasound (micro-US) offers real-time, high-resolution imaging for prostate cancer. Although artificial intelligence (AI) has shown potential in enhancing micro-US interpretation, a comprehensive review of this emerging field is currently missing. This review synthesizes current evidence on AI applied to ExactVu 29 MHz micro-US for prostate cancer. PubMed/MEDLINE, Embase, Scopus, Web of Science and the Cochrane Library were searched up to December 2025. Studies were included if they applied machine learning or deep learning directly to 29 MHz micro-US data and reported quantitative performance metrics. Ten studies met the inclusion criteria: six on prostate cancer detection, three on prostate segmentation and one on micro-US-histopathology registration. Detection models ranged from classical quantitative ultrasound machine learning to deep architectures using self-supervision, transformers, multiple-instance learning, ensemble calibration and 3D segmentation-based pipelines. Among core-level models for clinically significant cancer, area under the receiver operating characteristic curve (AUROC) values clustered around 0.76-0.81; one lesion-level framework reported an AUROC of 0.92, though at a non-comparable analytical unit. Segmentation studies achieved accurate prostate delineation (Dice similarity coefficient ≈ 0.94), and a single study demonstrated high-precision 3D registration to whole-mount histopathology (Dice similarity coefficient 0.97 and landmark error < 3 mm). All studies evaluated AI on previously acquired data, without real-time clinical implementation. AI for micro-US shows promising and reproducible early results across detection, segmentation and registration, but evidence is still limited. In view of the potential of AI to optimize micro-US utilization and its related advantages, additional efforts are warranted to achieve clinical adoption.