MRI, Prostate cancer diagnosis, and Artificial Intelligence; What are the implications?

Abstract

For many years now it has been suggested that artificial intelligence (AI) will come to have a significant role in healthcare delivery, although to date, the actual practical applications that are in everyday use have been limited in number. While the numbers of potential applications is increasing rapidly, the introduction of AI systems into healthcare is accompanied by a number concerns including issues of transparency, explainability, governance, bias, cybersecurity and quality assurance.

One area where AI has been thought to have significant potential is in the domain of image analysis, with potential applications in dermatology, histopathology and in radiology undergoing investigation. In recent months there have been two publications that provide evidence suggesting that AI has potential value in the interpretation of MRI scans in patients with suspected prostate cancer. These papers are not conclusive in themselves, but perhaps represent a step along the route to the widespread adoption of such systems into the diagnostic pathway. However, they also act as a practical example for articulating some of the general dilemmas that relate to the use of AI in healthcare.

The ability of the developed AI system was then compared with the diagnoses of 62 radiologists on a study population of 400 historical MRI scans, with the radiologists (who were experienced in reading prostate MRI) also having access to multiparametric imaging, although they did not have access to the patient history. In all cases, the ultimate histological diagnosis was known. The study design was “multi-reader multi-case” with the primary outcome being the diagnostic performance in terms of the identification of significant prostate cancer. The primary hypothesis was that AI was non-inferior to the radiologists, with an additional analysis against the “real-life” historical radiology reading that included multidisciplinary input. There was a secondary hypothesis that the AI system was superior to the radiologists.

The second study was undertaken in multiple different sites of the Mayo clinic using 5735 multiparametric MRI examinations in 5215 patients. There were 1514 clinically significant cases of prostate cancer. The AI system was trained on 5035 scans, with 300 scans used for validation and 400 examinations as the test set. The AI tool demonstrated an AUROC of 0.89 for both the AI tool and for the radiologists. The combination of the AI plus radiologist appeared to be superior to either individual approach.

It seems almost inevitable that AI will play an increasing role in modern healthcare, but while these publications highlight a potential role for an AI system in the interpretation of MRI scans of patients with potential prostate cancer, the associated issues outlined above will need to be carefully considered if such technology is to be introduced safely and effectively.

World News is written by Ian Eardley