An algorithm developed by American researchers has demonstrated the greatest precision to date in the recognition and characterization of prostate cancer, once again illustrating the advantages offered by artificial intelligence for this type of diagnosis.
” Humans are good at recognizing anomalies, but they have their own biases and past experiences “, valued Rajov Dhir, pathologist atUniversity of Pittsburgh (UPMC) and lead author of this study recently published in the journal The Lancet Digital Health. ” The machines do not suffer from such a bias, which certainly contributes to the standardization of the screening process.. “
To learn how toAI to identify the Prostate cancer, Dhir and his colleagues provided him with images of over a million stained tissue samples from biopsies. Each image has been tagged by expert pathologists to teach theAI distinguish healthy tissue from abnormal tissue. The algorithm was then tested on a separate set of 1,600 images of samples from 100 patients who were examined at the medical center of theUPMC for suspected prostate cancer.
During testing, theAI demonstrated a sensitivity of 98% and a selectivity of 97% in the detection of prostate cancer, a much higher precision than that previously achieved by algorithms working from tissue samples.
It is also the first algorithm to go beyond cancer detection, and show high performance for tumor classification, size, and spread to surrounding nerves, which are clinically important features. to be included in the pathology report.
“Such algorithms are particularly useful for atypical lesions”
TheAI also spotted six samples that had not been previously categorized by expert pathologists, but according to Dhir, this does not mean that the machine is necessarily superior to humans. During the evaluation of these cases, the pathologist may have simply seen enough evidence of malignancy in other samples from the patient to recommend treatment.
However, the algorithm could represent safety for less experienced pathologists, by detecting cases that might otherwise be missed.
” Such algorithms are particularly useful for atypical lesions, as a non-specialist might not be able to perform a correct assessment. “, highlighted Dhir. ” This is a major advantage of this kind of system. “
Although these results are particularly promising, the authors point out that new algorithms will have to be developed and trained to detect other types of cancer, since the pathological markers may differ depending on the types of tissues analyzed. However, researchers see no obstacle to adapting this type of technology to screen for other cancers, especially breast cancer.