Accurate identification of clinically significant prostate cancer remains a central challenge in urological imaging and biopsy planning. Conventional transrectal ultrasound-guided systematic biopsy is widely used but may miss important disease while detecting tumours that do not require treatment. Multiparametric magnetic resonance imaging (MRI) has improved detection and targeting, yet access, cost and variability in interpretation continue to influence its adoption. Quantitative multiparametric ultrasound combined with artificial intelligence offers an alternative approach by integrating multiple ultrasound-based measurements into a single analytical model. This strategy aims to provide objective tissue characterisation and support clinical decision-making before biopsy using widely available imaging technology.
Quantitative Multiparametric Ultrasound and Histology Integration
A prospective multicentre programme conducted in the Netherlands enrolled 604 patients between June 2021 and February 2024. After exclusion of incomplete or unusable datasets, 327 patients were analysed. All participants underwent transrectal three-dimensional multiparametric ultrasound imaging that included B-mode imaging, dynamic contrast-enhanced ultrasound and shear-wave elastography.
Must Read: AI for Personalised Breast Cancer Screening
The reference standard for clinically significant prostate cancer was derived from radical prostatectomy specimens. These specimens were processed into 4 mm histological slices, annotated and reconstructed into three-dimensional histology that was registered to the ultrasound imaging space. Clinically significant prostate cancer was defined as International Society of Urological Pathology (ISUP) Grade Group 2 or higher. Tissue within such lesions was labelled malignant, while other tissue was labelled benign. Malignancy probability between histological slices was estimated using interpolation, producing voxel-level probability maps used for model training and evaluation.
The dataset also included patients without clinically significant disease, confirmed through negative multiparametric MRI with PI-RADS 2 or lower and negative biopsy findings. This combination of imaging and histological reference data enabled detailed comparison between ultrasound-derived features and pathological outcomes.
Quantitative Imaging Features and Clinical Inputs
The analytical approach relied on extracting quantitative imaging features rather than visual interpretation. Dynamic contrast-enhanced ultrasound data were analysed using time–intensity curve modelling to describe vascular behaviour. Additional analysis estimated blood velocity and diffusion characteristics, as well as similarity between neighbouring vascular signals. Measures of variability in blood flow patterns were also calculated.
Shear-wave elastography contributed quantitative information about tissue stiffness. Three-dimensional elastography acquisition was performed through automated scanning of 25 planes from base to apex of the prostate. Elasticity and quality maps were generated and aligned with the histological reference data.
Clinical parameters were included alongside imaging features. Prostate volume was measured using automated segmentation of B-mode ultrasound images. Prostate-specific antigen density was calculated from prostate-specific antigen level and ultrasound-derived prostate volume. Together, these imaging and clinical variables were organised into three-dimensional feature maps describing tissue appearance, vascular dynamics, mechanical properties and clinical risk indicators.
Deep Learning Model Development and Validation
A deep learning model combining a three-dimensional convolutional neural network with a multilayer perceptron was trained to generate probability maps of clinically significant prostate cancer across the prostate gland. Training used seven-fold cross-validation on 250 patients. Because malignant tissue represented a small proportion of total voxels, undersampling techniques were applied during training to address class imbalance.
Uncertainty related to registration between histology and ultrasound, particularly near lesion boundaries, was incorporated into the training process through confidence-weighted loss functions. A control model using only anatomical location and zonal information was developed to assess the contribution of quantitative imaging features.
The full model achieved a voxel-wise area under the receiver operating characteristic curve of 0.870 during internal evaluation, with sensitivity of 76.8% and specificity of 80.2% at the optimal operating point. External validation used a temporally separated cohort of 77 patients enrolled after performance stabilised in the initial cohort. The model achieved an area under the receiver operating characteristic curve of 0.884, with no significant difference between internal and external performance.
The control model based on location alone achieved an area under the receiver operating characteristic curve of 0.730. Quantitative imaging feature groups improved performance beyond anatomical location alone, and combinations of feature groups produced stronger results across the whole prostate. Performance differed between anatomical zones, with higher values in the transitional zone than in the peripheral zone. Adding shear-wave elastography to contrast-enhanced ultrasound dispersion imaging features did not improve overall performance and slightly reduced performance in the transitional zone, reflecting limitations in elastography signal quality and sampling resolution.
Quantitative three-dimensional multiparametric ultrasound combined with deep learning enabled voxel-level identification of clinically significant prostate cancer with consistent performance in internal and external evaluation cohorts. The framework integrated quantitative biomarkers derived from ultrasound imaging together with prostate volume and prostate-specific antigen density, while accounting for class imbalance and registration uncertainty. Limitations included reliance on a single ultrasound system and contrast agent combination, a high proportion of radical prostatectomy cases in training data and the need for calibration across vendors. Rapid generation of dense probability maps suggests potential application in pre-biopsy triage and targeted biopsy workflows, with further development focused on lesion aggregation methods and comparison with MRI-targeted diagnostic pathways.
Source: European Radiology
Image Credit: iStock
