Abstract
Prostate cancer (PCa) lacks reliable and accurate tissue-based biomarkers to support prognostic stratification and clinical treatment decisions. Current diagnostic assessment, including Gleason grading, has limitations such as interobserver variability and insufficient granularity for disease aggressiveness. Fibroblast activation protein (FAP) and α-smooth muscle actin (αSMA) have emerged as putative stromal biomarkers, but their prognostic value in localised PCa has not been validated at scale. In this study, we developed a novel artificial intelligence (AI)-augmented image analysis pipeline tailored for dual-marker immunohistochemistry of FAP and αSMA, enabling automated, tissue compartment-specific quantification of biomarker expression. This deep learning model was trained and validated using digitised high-resolution whole-slide images of tissue microarrays from three prostatectomy cohorts, comprising 4,097 cores from 835 patients with comprehensive clinical follow-up data. The AI pipeline demonstrated high accuracy in detecting epithelial, stromal, and immune compartments, as well as in quantifying FAP and αSMA signals. We validated stromal FAP as a robust prognostic marker consistently associated with adverse clinical outcomes, including earlier biochemical recurrence, metastasis, and cancer-specific death. Epithelial FAP and stromal αSMA showed additional prognostic associations in selected analyses, particularly in MRI-visible tumours. Our findings reinforce the biological and clinical relevance of stromal FAP in the prostate tumour microenvironment. By enabling standardised and scalable biomarker quantification, our newly developed AI-assisted workflow advances the clinical utility of FAP and αSMA and demonstrates the power of integrating digital pathology with biomarker quantification. This study represents a critical step toward implementing stromal biomarkers in routine PCa diagnostics and underscores the potential of AI-enhanced histopathology in advancing precision oncology.