<p>Prostate cancer (PCa) is the second most diagnosed cancer among men and screening principally relies on standard clinical variables (PSA, prostate volume, PSAD, and age). Biparametric MRI (bpMRI) has emerged to support the visual assessment of clinically significant PCa (csPCa) lesions, improving PCa detection. Nevertheless, clinical decision-making requires malignancy stratification beyond csPCa likelihood estimations. Existing computer-based diagnostic models overlook the complementary information between clinical variables and bpMRI imaging, and only focus on estimating csPCa likelihood. This work proposes a deep learning network that integrates clinical variables with bpMRI representations to improve csPCa stratification into non-cancerous (NC), intermediate-malignant (IM), and high-malignant (HM) cases. A multimodal, geometric deep learning network is proposed, which effectively summarizes bpMRI information through symmetric positive definite (SPD) descriptors, addressing dataset imbalance and data scarcity. We developed an augmented clinical context input involving a digital radiologist model to encode expert radiological knowledge via PI-RADS scoring, image-derived anatomical ratios, categorical clinical vectors, and spatial information of analyzed regions. Multimodal imaging-variables representations were fused in an end-to-end manner with the proposed multi-stage fusion that hierarchically integrates clinical context at various levels of deep image representations, achieving more robust representations. In the unimodal imaging setup, the geometric approach improved F1-score from <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(80.9\pm 1.8\)</EquationSource> </InlineEquation> to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(88.9\pm 1.9\)</EquationSource> </InlineEquation>, while multi-stage fusion further increased performance to <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(90.2\pm 0.6\)</EquationSource> </InlineEquation>. For binary csPCa classification, precision improved to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(85.1\pm 4.3\)</EquationSource> </InlineEquation> compared with <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(80.1\pm 10.6\)</EquationSource> </InlineEquation> using the unimodal imaging model. The proposed method surpassed state-of-the-art methods in csPCa stratification and outperformed existing multimodal models in binary csPCa classification. This provides a potential clinical decision-support tool that improves diagnostic granularity and patient outcomes.</p>

A multimodal, geometric deep learning network integrating bpMRI and clinical variables for clinically significant prostate cancer stratification

期刊论文
发表日期：

2026年7月18日

查看全文

Juan A. Olmos,
Catalina de Valencia,
Antoine Manzanera,
Fabio Martínez

摘要

Prostate cancer (PCa) is the second most diagnosed cancer among men and screening principally relies on standard clinical variables (PSA, prostate volume, PSAD, and age). Biparametric MRI (bpMRI) has emerged to support the visual assessment of clinically significant PCa (csPCa) lesions, improving PCa detection. Nevertheless, clinical decision-making requires malignancy stratification beyond csPCa likelihood estimations. Existing computer-based diagnostic models overlook the complementary information between clinical variables and bpMRI imaging, and only focus on estimating csPCa likelihood. This work proposes a deep learning network that integrates clinical variables with bpMRI representations to improve csPCa stratification into non-cancerous (NC), intermediate-malignant (IM), and high-malignant (HM) cases. A multimodal, geometric deep learning network is proposed, which effectively summarizes bpMRI information through symmetric positive definite (SPD) descriptors, addressing dataset imbalance and data scarcity. We developed an augmented clinical context input involving a digital radiologist model to encode expert radiological knowledge via PI-RADS scoring, image-derived anatomical ratios, categorical clinical vectors, and spatial information of analyzed regions. Multimodal imaging-variables representations were fused in an end-to-end manner with the proposed multi-stage fusion that hierarchically integrates clinical context at various levels of deep image representations, achieving more robust representations. In the unimodal imaging setup, the geometric approach improved F1-score from \(80.9\pm 1.8\) to \(88.9\pm 1.9\) , while multi-stage fusion further increased performance to \(90.2\pm 0.6\) . For binary csPCa classification, precision improved to \(85.1\pm 4.3\) compared with \(80.1\pm 10.6\) using the unimodal imaging model. The proposed method surpassed state-of-the-art methods in csPCa stratification and outperformed existing multimodal models in binary csPCa classification. This provides a potential clinical decision-support tool that improves diagnostic granularity and patient outcomes.