Abstract
Algorithms matching the performance of expert pathologists in prostate cancer diagnosis were designed.
Major Finding: Algorithms matching the performance of expert pathologists in prostate cancer diagnosis were designed.
Concept: Deep learning was used to construct the algorithms, which also provide prognostic Gleason grades.
Impact: The use of such algorithms may reduce workload for pathologists and improve diagnostic accuracy.
The increasing number of prostate biopsies evaluated for cancer worldwide and the shortage of expert urological pathologists necessitate the development of tools to assist pathologists in diagnosis. In two separate studies, Bulten and colleagues and Ström, Kartasalo, and colleagues report the use of deep learning to develop algorithms based on computerized image scanning to diagnose prostate cancer from stained biopsy slides. When cancer was detected, each algorithm provided a Gleason grade—a strong correlate of disease aggressiveness and patient prognosis—and a tumor volume or length estimate. The algorithm presented by Bulton and colleagues outperformed pathologists with less than fifteen years of experience and its predictions did not differ significantly from those of a panel of expert pathologists, with some one-point errors observed in Gleason grading by the deep learning–based approach. Similarly, the approach presented by Ström, Kartasalo, and colleagues demonstrated performance within the range of experienced urological pathologists. Importantly, the exact accuracies of the two algorithms cannot be compared with one another because they were not validated using the same dataset or methods. For the same reason, it will be important to validate these findings in larger datasets from many different labs and scanners and to fine-tune the algorithms such that broader applicability is ensured. Although deep learning–based algorithms for diagnosis and grading of prostate cancer are not yet ready for clinical implementation, these studies present promising results suggesting that such a future is not out of reach. If successfully implemented, such algorithms could reduce the burden on pathologists by prescreening patient samples, improve diagnostic and prognostic accuracy by flagging samples for which the pathologist- and machine-derived results differ, and possibly even stand in for pathologists in areas where they are very scarce.
Note: Research Watch is written by Cancer Discovery editorial staff. Readers are encouraged to consult the original articles for full details. For more Research Watch, visit Cancer Discovery online at http://cancerdiscovery.aacrjournals.org/CDNews.
RSS Feeds