Machine Vision Poised to Take Another Bite Out of Human Docs' Sandwich

Computerized image analysis using rheumatoid arthritis patients' synovial samples was just about as good as human pathologists in quantifying joint inflammation, researchers said.

Sensitivity and specificity of the algorithm, with pathologists' assessments as the "gold standard," were 97% and 100%, respectively, in analyzing slides from 170 ordinary rheumatoid arthritis patients, according to Steven Guan, PhD, of the MITRE Corporation in McLean, Virginia, and colleagues.

The algorithm-generated results also correlated nicely with other disease activity measures including C-reactive protein, erythrocyte sedimentation rate, rheumatoid factor, and cyclized citrullinated protein antibodies, the researchers reported in ACR Open Rheumatology.

While robotic image analysis, a.k.a. machine vision, has sometimes been seen as a threat to pathologists' and radiologists' livelihoods, Guan and colleagues argued that this need not be the case.

In the pathology lab, they explained, the job is typically to pick out and characterize key features in histology slides. "By automating these fundamental tasks, the entire sample can be efficiently analyzed, and the burden on the pathologist and other human biases in selecting a subset of high-power fields from the whole slide image can be reduced," Guan's group wrote.

Yet they also indicated that automation would make the process more cost-effective, which is a backhand way of saying that humans cost too much when analyzing large numbers of samples. "Our site alone [the Hospital for Special Surgery in New York City, which supplied the patient samples and where many co-authors were based] performs over 5,000 arthroplasties per year," the group stated.

The algorithm they designed focused on counting nuclei in synovial tissue images "as a simple, fast, quantitative measurement of inflammation." Importantly, a single slide may have "hundreds of thousands" of nuclei, Guan and colleagues explained, such that exact counts are beyond the reach of human pathologists.

Tissue samples for the study were obtained from rheumatoid arthritis patients undergoing arthroplasty, so the entire synovium was available for analysis. Tissues were stained with hematoxylin and eosin (H&E) and placed on slides, which were then photographed in the usual manner. These were then subjected to the machine-vision analysis.

Experienced pathologists then checked a sampling of 10 images to determine whether the algorithm had correctly identified nuclei. These analysts also performed histologic scoring on samples, focusing on areas that seemed to be the most inflamed. Gene expression and standard lab results from patients were examined as well.

A median of about 112,000 nuclei were counted per sample by the algorithm. Guan and colleagues indicated that a range of H&E concentrations were applied to the samples, and the variation did not seem to affect the algorithm's performance.

Another aim of the study was to confirm that nuclei density is a useful index of joint inflammation, and it did, insofar as it correlated strongly with other, conventional measures. These included histologic score, expression of inflammation-associated genes, and patients' customary lab values.

But that wasn't true for many standard measures, including disease duration, swollen and tender joint counts, pain, and patients' self-assessments, for which no significant correlation was seen.

Why that was the case remains unclear, but the researchers pointed to the likelihood that, given that patients were undergoing arthroplasty, they had what amounts to "end-stage" joint disease and therefore aren't reflective of the overall rheumatoid arthritis population.

Guan and colleagues cautioned, too, that the algorithm isn't yet ready for widespread use, calling the current work "an initial step" toward a sufficiently reliable and robust product that physicians can use confidently.

Comment