COPENHAGEN — Software that relies on the power of artificial intelligence to "teach" itself can distinguish between prostate samples that are cancerous and those that are not, and could soon be used to determine a Gleason score for the sample, new research shows.
"The system was programmed to learn and to gradually improve how it interprets the samples," investigator Hongqian Guo, MD, from Nanjing Drum Tower Hospital in China, explained in a statement.
"Our results show that the diagnoses the AI reported was at a level comparable to that of a pathologist," he reported during a poster session here at the European Association of Urology 2018 Congress. And "it could accurately classify the malignant levels of prostate cancer."
Guo and his colleagues used the smart software system, which was developed in conjunction with Nanjing Innovative Data Technologies, Inc., to assess 918 whole-mount prostate samples from 283 patients who had undergone radical prostatectomy. The evaluation was made in accordance with the 2014 International Society of Urological Pathology grading system for prostate cancer (Am J Surg Pathol. 2016;40:244-252).
The team then identified regions of different tumor grades within the whole-mount section, from low (grades 1 and 2), to intermediate (grades 3 and 4), to high (grade 5).
It took a considerable amount of time for the machine to learn how to distinguish cancerous from noncancerous samples, but the diagnostic accuracy of the software gradually improved over time, said investigator Chengwei Zhang, MD, PhD, also from the Nanjing Drum Tower Hospital.
They then tested the ability of the software to distinguish whether a cancer was present or not by giving the machine 10 pieces of whole-mount sections from 10 different patients. It was accurate 99.3% of the time.
And when they compared tumor grades determined by the software with those determined by a pathologist, they found that the results corresponded well.
"We believe this is the first automated work to offer an accurate reporting and diagnosis of prostate cancer," said Guo.
More to Learn
The machine still has to learn more, especially how to grade lower- and higher-grade Gleason score cancers. Most of the samples it has been exposed to have been intermediate-risk prostate cancers, Zhang pointed out.
With continued improvement, the software might eventually be able to diagnose not just prostate cancer, but other cancers as well.
It is also possible that the software will become better at diagnosing prostate cancer and its level of malignancy than an inexperienced pathologist, who in turn could learn from the software how to better identify samples that contain significant disease, the investigators explain.
"This is not going to replace a human pathologist," Guo said. "We still need an experienced pathologist to take responsibility for the final diagnosis."
But it will "help pathologists make better, faster diagnoses" and "eliminate day-to-day variations in judgment that can creep into human evaluations," he added.
"This may be very useful in some areas where there is a lack of trained pathologists," Rodolfo Montironi, MD, from the Marche Polytechnic University in Ancona, Italy, said.
However, Montironi cautioned, even though smart software will diminish the reliance on human expertise, "we still need to ensure that the final decisions on treatment stay with a trained pathologist."
Guo, Zhang, and Montironi have disclosed no relevant financial relationships.
European Association of Urology (EAU) 2018 Congress: Abstract 213. Presented March 16, 2018.
Follow Medscape on Twitter @Medscape
Medscape Medical News © 2018 WebMD, LLC
Send comments and news tips to email@example.com.
Cite this: System Gaining in Diagnostic Duel Between Doctor and Machine - Medscape - Mar 21, 2018.