Comprehensive Summary
This study developed a foundation model, MRI-PTPCa, trained on both MRI and paired pathology data to produce a noninvasive diagnosis and grading of prostate cancer. Using approximately 1.3 million image-pathology pairs from a population of over 5,500 patients, the model was trained and its predictions in real-world settings had a grading accuracy of approximately 89.1%. In comparison to current invasive methods, the estimations produced by MRI-PTPCa were superior. Being a user-friendly tool, the MRI-PTPCa model allows for the user to more seamlessly bridge the gap between radiology and pathology in a precise manner and reduces diagnostic ambiguities in prostate cancer gray zones. Due to the model’s emphasis on improving the specificity (SPE) of the prostate-specific antigen (PSA), a molecule that has been demonstrated to be associated with aggressive prostate cancer, without increasing the sensitivity (SEN) of the Prostate Imaging Reporting and Data System (PI-RADs), the model was found to decrease the number of overdiagnosed CSPCa cases by approximately 90%, reducing the instances of incorrect treatment and the resource burden caused by overuse of SEN
Outcomes and Implications
This study holds numerous clinical implications in prostate cancer. Not only is it a non-invasive method for diagnosis, but the accuracy of the MRI-PTPCa model demonstrates a more reliable, precise, and consistent mechanism for diagnosis. Through a more comprehensive diagnosis, the risk for over and under treatment is minimized, resulting in a more effective degree of clinical management of prostate cancer. Due to its accuracy and format, it also may reduce the resource burden in clinical settings by minimizing the need for additional screening. Additionally, the foundation of this model may be expanded to address other forms of under or over diagnosed cancers in the future, enabling healthcare providers to most accurately diagnose patients and address treatment methods.