PiCaRad: Multi-modal MRI Radiomics Dataset for Whole-gland and Cancer Lesion Characterisation

Abstract

Abstract Purpose: To develop PiCaRad, a standardized radiomics-ready prostate MRI resource that facilitates reproducible radiomics research by providing analysis-ready feature matrices, harmonized clinical annotations, and reference benchmark experiments. Methods: PiCaRad was constructed from the publicly available PI-CAI and Prostate158 datasets using a unified PyRadiomics pipeline. Standardized whole-gland and lesion radiomic features were extracted from T2-weighted (T2W), apparent diffusion coefficient (ADC), and high b-value diffusion-weighted (HBV) MRI following quality-controlled segmentation and harmonized preprocessing. The resource integrates clinical variables, acquisition-domain metadata, segmentation agreement metrics, benchmark machine learning models, and domain shift analyses. Whole-gland, lesion-level, clinical, multimodal, unified multi-dataset, and external validation experiments were performed using standardized machine learning pipelines. Results: PiCaRad comprises standardized whole-gland radiomic feature matrices from 1,204 PI-CAI examinations (T2W, ADC, and HBV), 178 expert-derived lesion radiomics cases, 321 AI-derived lesion radiomics cases, and 158 standardized whole-gland T2W radiomics cases from the independent Prostate158 dataset. The resource additionally includes harmonized clinical metadata, acquisition-domain annotations, segmentation quality assessment results, and reproducible benchmark machine learning experiments. Conclusion: PiCaRad provides a comprehensive, publicly available radiomics-ready prostate MRI resource that reduces preprocessing effort and establishes standardized benchmark experiments for reproducible radiomics, machine learning, and future multimodal prostate cancer research.