AI Can Accurately Assign Contrast Use in Prostate MRI

Training sets from different vendors may be required to ensure scanner-specific sensitivity.

An artificial intelligence (AI) algorithm was able to independently decide on the necessity of contrast agent application in prostate magnetic resonance imaging (MRI) with high accuracy and a very low false negative rate.

Detailed in a study published in Insights into Imaging, Dr. Andreas Hötker, of the Institute of Diagnostic and Interventional Radiology at University Hospital Zurich in Switzerland, and colleagues, developed and validated a convolutional neural network (CNN) to decide on the necessity of dynamic contrast-enhanced (DCE) sequences in prostate MRI.

“The CNN would have correctly assigned 78% of patients to a biparametric or multiparametric protocol, with only 2% of all patients requiring re-examination to add DCE sequences,” the authors said. “Integrating this CNN in clinical routine could render the requirement for on-table monitoring obsolete by performing contrast-enhanced MRI only when needed.”

Multiparametric prostate MRI is part of routine practice in detecting and staging prostate cancer, but several studies have reported a comparable performance of an abbreviated MRI protocol without the use of a contrast agent. Along with saving time, biparametric prostate MRI avoids any contrast agent side effects, improves cost-effectiveness and optimizes radiology workflow. However, multiparametric MRI reduces the number of indeterminate lesions and is valuable in examinations with poor image quality or for the less-experienced radiologists.

For more expert insights and news, subscribe to the Diagnostic Imaging e-Newsletter here.

While the decision on whether to perform DCE should ideally be made on a per-scan basis, applying on-table monitoring and having a radiologist make this ad-hoc based decision is not often feasible in routine practice. Hötker and colleagues sought to develop and validate a CNN that would allow for real-time decision-making, automatically identifying patients who would benefit from acquisition of a DCE sequence.

The CNN was developed on 300 prostate MRI examinations and validated in a separate cohort of 131 prostate MRI examinations. The CNN reached a sensitivity of 94.4% and specificity of 68.8% for DCE necessity, correctly assigning 44% of patients to a biparametric protocol and 34% to a multiparametric protocol. The CNN incorrectly decided on omitting DCE in 2% of patients. Meanwhile, the decisions made by a radiology technician had a sensitivity of 63.9% and specificity of 89.1%.

When the CNN was applied to a set of MRI examinations performed on a different scanner, a sensitivity of 100% was achieved, but with a specificity of 42.1%. A dedicated training set based on MRI scans from this different vendor would likely increase the accuracy of the CNN, the authors suggested.

“Integration of AI into quality assessment and decision making could allow for shorter examination times and a more streamlined clinical workflow, while maintaining diagnostic accuracy by including DCE only when truly needed,” the authors said.