Can Generative AI Facilitate Simulated Contrast Enhancement for Prostate MRI?

New research showed that simulated contrast-enhanced MRI provided a high multiscale structural similarity index and robust inter-reader agreement with contrast-enhanced MRI in diagnosing clinically significant prostate cancer.

For the retrospective study, recently published in Radiology, researchers assessed a deep learning algorithm’s use of a conditional generative adversarial network (CGAN) to provide simulated contrast-enhanced MRI derived from four non-contrast multiparametric MRI (mpMRI) sequences for T1 and T2-weighted MRI, diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC). The study authors compared the simulated contrast-enhanced MRI scans to actual contrast-enhanced MRI in 567 patients (mean age of 66) with suspected PCa.

The researchers found that simulated contrast-enhanced MRI demonstrated an average 70 percent multiscale structural similarity index with actual contrast-enhanced MRI in two sets of external validation testing. There was also a 96 percent Cohen k assessment for inter-reader agreement between three reviewing radiologists on PI-RADS scoring based on T2-weighted MRI and DWI sequences, according to the study.

“Excellent concordance of lesion conspicuity in the peripheral zone and transition zone was observed between acquired and simulated contrast-enhanced imaging, even in patients with insufficient DWI quality,” wrote lead study author Hongyan Huang, M.D., who is affiliated with the Department of Radiology at Shenzhen Nanshan People’s Hospital and Shenzhen University in Guangdong, China, and colleagues.

The study authors added that changes from initial PI-RADS 3 assessments to PI-RADS 4 scoring occurred in 10.5 percent of patients after the addition of simulated contrast-enhanced MRI scans to biparametric MRI.

“This score change may have increased the depiction of clinically significant prostate cancer; reduced unnecessary biopsies; or enabled additional targeted biopsy, which is important for management decisions in patients who have not undergone biopsy, particularly given the trend toward avoiding biopsy in patients with low-risk PI-RADS 3 scores,” emphasized Huang and colleagues.

(Editor’s note: For additional content on prostate cancer imaging, click here.)

In an accompanying editorial, Radhouene Neji, Ph.D., and Vicky Goh, M.D., said prospective multicenter studies are necessary for further illumination of the study findings, and questioned whether the input data for generative AI models would have sufficient clinical and/or physiologic information to facilitate simulated contrast-enhanced MRI.

However, noting that contrast-enhanced MRI is not considered a primary defining sequence for PI-RADS scoring, Drs. Neji and Goh suggested that generative AI may have merit in this context.

“Omitting intravenous contrast material may reduce scanning time, may be more cost effective, and may allow more male patients access to MRI, an important consideration with the projected incidence of prostate cancer. Amid ongoing debate and with results of prospective trials yet to be published, research into generative artificial intelligence (AI) to synthesize contrast-enhanced MRI scans is timely and relevant,” noted Dr. Neji, a senior lecturer in magnetic resonance physics and computing at the School of Biomedical Engineering and Imaging Sciences at King’s College in London, and Dr. Goh, a professor and head of the Department of Cancer Imaging at the aforementioned institution.

(Editor’s note: For related content, see “Could MRI-Based AI Offer Better Risk Stratification for Prostate Cancer than PI-RADS?,” “MRI Study Suggests Deep Learning Model Offers Equivalent Detection of csPCa as Experienced Radiologists” and “MRI-Based AI Model Facilitates 50 Percent Reduction in False Positives for Prostate Cancer.”)

In regard to study limitations, the study authors conceded potential patient selection bias with the retrospective nature of the research. They pointed out that bowel and smooth muscle movements may have comprised co-registration among MRI sequences and added that the requirement of T1-weighted imaging for simulated contrast enhancement may adversely affect broader application of the model.