AI on par with radiologists for finding prostate cancer on MRI

AI may be on par with radiologists for finding prostate cancer detection on MRI, according to research published October 13 in Radiology. 

An AI system achieved a diagnostic performance comparable to that of radiologists for finding Gleason grade group ≥2 prostate cancer across external diagnostic and screening datasets, wrote a team led by Anindo Saha, PhD, from Radboud University Medical Center in Nijmegen, the Netherlands. 

“These findings support a potential role for AI in extending prostate MRI interpretation, particularly where subspecialty expertise is limited,” the Saha team wrote. 

MRI is the go-to method for diagnosing prostate cancer, but barriers to accessing exams prevent equitable care. This especially goes for areas with radiologist shortages as demand rises for medical imaging. 

AI has shown potential for addressing these concerns, with prior research suggesting the technology can achieve non-inferior performance compared to radiologist readings. Saha and colleagues noted that large-scale external evaluations of AI for prostate cancer detection are limited.  

The Scaling Artificial Intelligence for Prostate Cancer Detection on MRI to Global, Multiethnic Populations (SCARLET-1) is one study that addressed these gaps, taking place across 46 cities in 22 countries. It determined whether an AI system can achieve diagnostic performance comparable to that of radiology readings for the detection of Gleason grade group ≥2 prostate cancer on MRI in external diagnostic and screening datasets. The study also evaluated factors that can influence whether AI systems can maintain reliability within local clinical workflows. 

Prostate MRI exams from the internal test sets, illustrating AI detection of Gleason grade group ≥2 cancers across diverse international sites, scanners, and MRI appearances. From top-to-bottom, the exams were performed at clinical sites based in Hanoi, Vietnam; Chicago, United States; Nairobi, Kenya; and Amsterdam, the Netherlands. Columns show, from left-to-right, the AI detection (if any) overlaid on the T2-weighted image, the T2-weighted image alone, the high-b-value diffusion-weighted image, and the apparent diffusion coefficient map. Colored contours represent regions identified by the AI system as suspicious for Gleason grade group ≥2 prostate cancer; exams without a colored overlay had no AI-detected region. The overlay color represents the AI suspicion score on a continuous scale from 0–100%, with green, orange, and red overlays in the examples shown corresponding to scores of 55%, 80% and 93%, respectively. All images are of the axial plane and have been cropped to show the region of interest. Intravenous gadolinium-based contrast material was administered for the exams; however, dynamic contrast-enhanced MRI sequences were not included in the study analyses. The male patient characteristics by row are as follows: Hanoi–75 years of age, PSA 100 ng/mL, PI-RADS 5, Gleason grade group 4; Chicago—57 years of age, PSA 12 ng/mL, PI-RADS 4, Gleason grade group 5; Nairobi—59 years of age, PSA 17 ng/mL, PI-RADS 2, Gleason grade group 2; and Amsterdam—67 years of age, PSA 7 ng/mL, PI-RADS 3, no cancer found on biopsy. Imaging was performed using a 1.5-T SIGNA Creator (GE HealthCare), a 3.0-T Skyra (Siemens Healthineers), a 1.5-T SIGNA Explorer (GE HealthCare), and a 3.0-T Achieva dStream (Philips Medical Systems), respectively. A urinary catheter is present in the exam from Hanoi, and benign prostatic hyperplasia is present in the exam from Chicago and Amsterdam.RSNA

Saha and colleagues reported their findings from SCARLET-1, which included 22,445 prostate MRI exams from 21,252 men. The men underwent MRI scans between 2012 and 2024. Of the total exams, the team used 20,471 from 19,278 patients for AI training and internal testing, and another 1,974 patients for external testing.  

In both diagnostic and screening settings, the AI system achieved high performance and high agreement with radiology readings for diagnosing Gleason grade group ≥2 cancer. 

The researchers also reported consistent performance in subgroup analysis. This included factors such as age, risk, cancer location, and ethnicity (mean ΔAUC, 0.02). However, performance varied across 25 sites, including 28% with AUC < 0.80 and 32% with AUC ≥ 0.90). Image quality also varied, with AUCs ranging between 0.78 and 0.88. 

These secondary findings show how AI performance is dependent on clinical context, the researchers wrote, adding that performance varies with case mix and image quality across datasets. 

The study authors called for further large-scale trials to provide complementary evidence. 

“Safe clinical use will require careful operating point selection, local testing, and longitudinal performance monitoring, and prospective evaluation of clinical, equity, and cost outcomes,” they added. 

Read the full study here.