Automated system grades clear cell RCC samples
BY NEIL Osterweil
Renal pathology is both a science and an art, requiring a breadth of knowledge about normal and abnormal tissues and a sharp, discerning eye that can detect patterns and features that can signal the presence and extent of malignancy.
Pathologists, with appropriate training and extensive experience, excel at analyzing complex visual patterns, but opinions about the gray areas – tissues with uncertain or borderline morphology – can differ and lead to different diagnostic conclusions.
A team of investigators in Singapore is attempting to take some of the guesswork out of renal histopathology for more accurate and more consistent assessments using an automated system designed to recognize nuclear pleomorphic patterns and analyze prominent nucleoli for grading of clear cell renal cell carcinoma slides.
“The Fuhrman grading system predominantly focuses on the morphology of the nucleus and the existence of prominent nucleoli. We attempted to capture some of this information and provide a quantitative assessment of patient histopathologic images,” wrote Daniel Aitor Holdbrook, PhD, of the Bioinformatics Institute in Singapore and his colleagues.
“Critically, we have demonstrated that, by focusing on prominent nucleoli, we are able to achieve high predictive power, which indicates that this feature and the features of the surrounding nucleus play a dominant role in cancer grading,” they reported in a study published in JCO Clinical Cancer Informatics.
The investigators retrospectively collected histopathology slides from 59 patients who underwent surgery for clear cell renal cell carcinoma at Singapore General Hospital.
They ran the slides through what they describe as an “automated image classification pipeline” that grades images from whole pathology slides.
“This pipeline uses two separate prominent nucleoli detectors to create two sets of prominent nucleoli image patches from a given image. The pipeline then quantifies nuclear pleomorphic patterns by concatenating features extracted from multiple prominent nucleoli image patches,” the investigators explained.
The concatenated, or linked, image “patches” are then analyzed and a fraction value (FV) score is generated. This score determines the fraction of features predicted to be low or high in a given detection set for both sets of prominent nucleoli image patches.
Finally, the system uses the two fraction value scores to grade the images as either low (Fuhrman grade 1 or 2) or high (Fuhrman grade 3 or 4).
The investigators compared the FV score with a validated multigene assay score obtained from 62 kidney samples from 58 patients, and they determined a correlation of 0.59 between the average FV predicted for a region in a slide, and the multigene score for the corresponding tissue in the same region.
“This high degree of correlation further validates the automated image-based classification pipeline and provides a molecular link to nuclear-based image grading. The multigene score has already been demonstrated to be highly correlated with patient prognosis, and its correlation with FV score further validates the use of this measure in grading histopathologic images,” Dr. Holdbrook and his associates wrote.
No substitute (yet) for the human eye
But a renal cancer expert who was not involved in the study said that, for the moment at least, renal pathologists need not fear being made redundant by their robot overlords.
“The first question that comes up is what is the value of an automated system vs. the human eye? And I think that, at least in this day and age, an expert human eye in general still does better than these automated algorithms,” said Walter M. Stadler, MD, Fred C. Buffett Professor of Medicine and Surgery and the chief of the section of hematology/oncology at the University of Chicago.
He cautioned that the operative word in that statement is “expert” and that it only applies to pathologists with extensive training and experience in interpreting renal carcinoma slides, with the assumption that the pathologist has sufficient time to analyze and grade tissues.
Nonetheless, he said, “I think that research like this is really valuable, and that in the long run we will get to a place where using these kinds of algorithms will much more quickly and just as accurately as a bona fide expert come up with grades and nuclear scoring.”
Dr. Stadler said that the necessary next step for Dr. Holdbrook and his associates will be to validate the findings in larger, preferably prospectively-collected data sets.
SOURCE: Holdbrook DA et al. JCO Clin Cancer Inform. 2018 Apr 16. doi: 10.1200/CCI.17.00100.