Artificial Intelligence and Radiologists at Prostate Cancer Detection in MRI: The PI-CAI Challenge

Overview

The PI-CAI challenge aims to validate the diagnostic performance of artificial intelligence (AI) and radiologists at clinically significant prostate cancer (csPCa) detection/diagnosis in MRI, with respect to histopathology and follow-up (≥ 3 years) as reference. The study hypothesizes that state-of-the-art AI algorithms, trained using thousands of patient exams, are non-inferior to radiologists reading bpMRI. As secondary end-points, it investigates the optimal AI model for csPCa detection/diagnosis, and the effects of dynamic contrast-enhanced imaging and reader experience on diagnostic accuracy and inter-reader variability.

Prostate cancer (PCa) is one of the most prevalent cancers in men worldwide. One million men receive a diagnosis and 300,000 die from clinically significant PCa (csPCa) (defined as ISUP≥2), each year, worldwide. Multiparametric magnetic resonance imaging (mpMRI) is playing an increasingly important role in the early diagnosis of prostate cancer, and has been recommended by the European Association of Urology (EAU), prior to biopsies. However, current guidelines for reading prostate mpMRI (i.e. PI-RADS v2.1) follow a semi-quantitative assessment that mandates substantial expertise for proper usage. This can lead to low inter-reader agreement (\<50%), sub-optimal interpretation and overdiagnosis. Modern artificial intelligence (AI) algorithms have paved the way for powerful computer-aided detection and diagnosis (CAD) systems that rival human performance in medical image analysis. Clinical trials are the gold standard for assessing new medications and interventions in a controlled and comparative manner, and the equivalent for developing AI algorithms are international competitions or "grand challenges", where increasingly large datasets are released to public to solve clinically relevant tasks with AI. Grand challenges can address the lack of trust, scientific evidence and adequate validation among AI solutions, by providing the means to compare algorithms against each other using common datasets and a unified experimental setup. PI-CAI (Prostate Imaging: Cancer AI) is an all-new grand challenge, with over 10,000 carefully-curated prostate MRI exams to validate modern AI algorithms and estimate radiologists' performance at csPCa detection and diagnosis. Key aspects of the study design have been established in conjunction with an international, multi-disciplinary scientific advisory board (16 experts in prostate AI, radiology and urology) -to unify and standardize present-day guidelines, and to ensure meaningful validation of prostate-AI towards clinical translation. The 2022 edition of PI-CAI will focus on validating AI at automated 3D detection and diagnosis of csPCa in bpMRI. PI-CAI primarily consists of two sub-studies: * AI Study (Grand Challenge): An annotated multi-center, multi-vendor dataset of 1500 bpMRI exams (including their clinical and acquisition variables) is made publicly available for all participating teams and the research community at large. Teams can use this dataset to develop AI models, and submit their trained algorithms (in Docker containers) for evaluation. At the end of this open development phase, all algorithms are ranked, based on their performance on a hidden testing cohort of 1000 unseen scans. In the closed testing phase, organizers retrain the top-ranking 5 AI algorithms using a larger dataset of 7500-9500 bpMRI scans (including additional training scans from a private dataset). Finally, their performance is re-evaluated on the hidden testing cohort (with rigorous statistical analyses), to determine the top 3 AI algorithms for automated 3D detection and diagnosis of csPCa in bpMRI (i.e. the winners of the grand challenge). * Reader Study: 50+ international prostate radiologists perform a reader study using a subset of 400 scans from the hidden testing cohort. For each case, radiologists complete their assessments in two rounds. At first, using clinical and acquisition variables plus bpMRI sequences only, enabling head-to-head comparisons against AI trained on the same. And then, using clinical and acquisition variables plus full mpMRI sequences, enabling comparisons between AI and current clinical practice (PI-RADS v2.1). Overall, the goal of this study is to estimate the performance of the average radiologist at detection and diagnosis of csPCa in MRI. In the end, PI-CAI aims to benchmark state-of-the-art AI algorithms developed in the grand challenge, against prostate radiologists participating in the reader study -to evaluate the clinical viability of modern prostate-AI solutions at csPCa detection and diagnosis in MRI.